CoChat
Sign Up
Cited by 2,550Psychiatry, Mental Health, NeuroscienceAI-based Problem Solving and PlanningCognitive Computing and Networks

Building machines that learn and think like people

Open access
Lake, Brenden M., Ullman, Tomer D., Tenenbaum, Joshua B., Gershman, Samuel J.·2017·Behavioral and Brain Sciences
DOI 10.1017/s0140525x16001837
Open this paper in CoChat — free accountOpen PDF

Abstract

Recent progress in artificial intelligence has renewed interest in building systems that learn and think like people. Many advances have come from using deep neural networks trained end-to-end in tasks such as object recognition, video games, and board games, achieving performance that equals or even beats that of humans in some respects. Despite their biological inspiration and performance achievements, these systems differ from human intelligence in crucial ways. We review progress in cognitive science suggesting that truly human-like learning and thinking machines will have to reach beyond current engineering trends in both what they learn and how they learn it. Specifically, we argue that these machines should (1) build causal models of the world that support explanation and understanding, rather than merely solving pattern recognition problems; (2) ground learning in intuitive theories of physics and psychology to support and enrich the knowledge that is learned; and (3) harness compositionality and learning-to-learn to rapidly acquire and generalize knowledge to new tasks and situations. We suggest concrete challenges and promising routes toward these goals that can combine the strengths of recent neural network advances with more structured cognitive models.

AI snapshot

Written from a shortened version of the paper’s full text (it is longer than we read).

TL;DR

The paper argues that deep learning systems can match or exceed human performance on some pattern-recognition tasks but still lack key ingredients of human-like learning and thought, especially causal model building, intuitive physics and psychology, compositionality, and learning-to-learn. It uses handwritten character learning and the Atari game Frostbite as challenge problems to show where current systems fall short and to propose routes toward more human-like machines.

Key findings

  • The DQN needed about 924 hours of unique experience per Atari game, roughly 500 times the approximately 2 hours given to a professional gamer, and achieved less than 10% of human-level Frostbite performance in a controlled test; later variants reached 83% and 172% of the expert score but still required far more experience. In the authors’ informal play, after watching about 2 minutes of expert video, they reached expert-comparable Frostbite scores within 15–20 minutes of practice.
  • On handwritten characters, deep convolutional networks approach human error on MNIST at 0.2%, but one-shot learning differs sharply: Bayesian program learning achieved human-level one-shot classification and generation after pretraining on five alphabets (about 150 character types), whereas deep ConvNets had 23% error versus 4% for humans with five alphabets, and 30% error with 30 alphabets.
  • AlphaGo was initially trained on 28.4 million positions from 160,000 human games and then improved through 30 million self-play games, while Lee Sedol is estimated to have played around 50,000 games in his life; the paper uses this contrast to argue that human-like understanding and flexible generalization remain distinct from benchmark performance.
  • The paper proposes core ingredients for human-like intelligence: intuitive physics, intuitive psychology, compositionality, causality, learning-to-learn, and fast approximate inference or model-free support for slower model-based reasoning.

Methods

  • Conceptual review and synthesis drawing on cognitive science, developmental psychology, and artificial intelligence; no new formal experiment is reported.
  • Two running challenge problems are used: the Characters Challenge for one-shot handwritten character concept learning, and the Frostbite Challenge for deep reinforcement learning control.
  • The analysis contrasts pattern recognition with model building, reviews evidence on human learning and machine learning, and proposes integration routes; it also includes informal author play of Frostbite after watching expert videos.

Limitations

  • Stated: the ingredient list is not intended to be complete; language is acknowledged as essential but not developed; the authors remain agnostic about whether the ingredients are innate or learned; and the focus is on neural networks, with probabilistic machine learning largely beyond scope.
  • Evident: there is no controlled empirical study; the Frostbite human comparison relies on informal author play and prior experience; and the challenge comparisons may be confounded by humans’ extensive prior knowledge, which the authors acknowledge but do not fully resolve.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1017/s0140525x16001837

Cites

274 of 274 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 393 cited sources checked · 0 supported · 0 flagged

  • Findingnone given

    After a small number of games over minutes, humans can understand Frostbite and its goals well enough to outperform deep networks that have almost a thousand hours of experience.

    “Humans, after playing just a small number of games over a span of minutes, can understand the game and its goals well enough to perform better than deep networks do after almost a thousand hours of experience”
    Quote in the paperNot a substantial claim
  • Findingown data

    Watching an expert play for a few minutes lets humans learn Frostbite even faster.

    “If humans are able to first watch an expert playing for a few minutes, they can learn even faster”
    Quote in the paperNot a substantial claim
  • Findingreasoning

    A trained DQN network is inflexible to changes in its inputs and goals.

    “The learned DQN network is also rather inflexible to changes in its inputs and goals”
    Quote in the paperSubstantial claim
  • Findingnone given

    Random play on Frostbite achieves a score of 65.2.

    “Random play achieves a score of 65.2”
    Quote in the paperNot a substantial claim
  • Findingown data

    In informal experiments, after two minutes of watching expert Frostbite videos, two authors reached scores comparable to or better than the human expert reported by Mnih et al. (2015) after at most 15 to 20 minutes of practice.

    “after watching videos of expert play on YouTube for just 2 minutes, we found that we were able to reach scores comparable to or better than the human expert reported in Mnih et al. (2015) after at most 15 to 20 minutes of total practice”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Mnih et al. (2015)

      Not checked in the free run.

  • May combine two claims. 156
  • Classified as probably not a substantial claim. 166
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingnone given

    Even the best deep networks learn gradually over many thousands of game episodes, take a long time to reach good performance, and are locked into particular input and goal patterns.

    “Even the best deep networks learn gradually over many thousands of game episodes, take a long time to reach good performance, and are locked into particular input and goal patterns”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingcitation

    Probabilistic program induction has made progress on the Characters Challenge, but aspects of full human cognitive ability remain out of reach.

    “We recently reported progress on this challenge using probabilistic program induction (Lake et al. 2015a) (see Glossary in Table 1), yet aspects of the full human cognitive ability remain out of reach.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Lake et al. 2015a

      Not checked in the free run.

  • May combine two claims.

    Findingcitation

    Human and PhysNet confidence were correlated across towers, but less strongly than for the approximate probabilistic simulation models and experiments of Battaglia et al. (2013).

    “Human and PhysNet confidence were also correlated across towers, although not as strongly as for the approximate probabilistic simulation models and experiments of Battaglia et al. (2013)”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Lerer et al. 2016

      Not checked in the free run.

    • cites Battaglia et al. 2013

      Not checked in the free run.

  • May combine two claims.

    Findingcitation

    PhysNet generalized to simple real images of block towers, matching human performance on those images and exceeding human performance on synthetic images.

    “Impressively, PhysNet generalized to simple real images of block towers, matching human performance on these images, meanwhile exceeding human performance on synthetic images”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Lerer et al. 2016

      Not checked in the free run.

  • May combine two claims.

    Findingcitation

    Causal knowledge influences how people learn new concepts; providing different types of causal knowledge changes how they learn and generalize.

    “Causal knowledge has also been shown to influence how people learn new concepts; providing a learner with different types of causal knowledge changes how he or she learns and generalizes.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation

    The Bayesian Program Learning (BPL) framework performs one-shot classification at human-level performance and outperforms current deep learning models such as convolutional networks.

    “Learning models of this form allows BPL to perform a challenging one-shot classification task at human-level performance (Fig. 1A-i) and to outperform current deep learning models such as convolutional networks (Koch et al. 2015).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Koch et al. 2015

      Not checked in the free run.

  • May combine two claims.

    Findingcitation

    Deep neural network image captioning often gets key objects correct but fails to understand physical forces, mental states, or causal relationships.

    “In many cases, the network gets the key objects in a scene correct, but fails to understand the physical forces at work, the mental states of the people, or the causal relationships between the objects.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation

    Graves's model synthesizes impressive handwriting in various styles but requires a large training corpus and has not been applied to other tasks.

    “Although it synthesizes impressive examples of handwriting in various styles, it requires a large training corpus and has not been applied to other tasks.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Graves 2014

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Findingnone given

    Humans can still reach these scores with a few minutes of practice, requiring far less experience than the DQNs.

    “But anecdotally, we find that humans can still reach these scores with a few minutes of practice, requiring far less experience than the DQNs.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given

    Transfer learning and multitask learning have not yet led to systems that learn new tasks as rapidly and flexibly as humans do.

    “Although transfer learning and multitask learning are already important themes across AI, and in deep learning in particular, they have not yet led to systems that learn new tasks as rapidly and flexibly as humans do.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    With this level of pre-training, current neural networks perform much worse on classification and have not even attempted generation.

    “With this level of pre-training, current neural networks perform much worse on classification and have not even attempted generation; they are still far from solving the Characters Challenge.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation

    Deep reinforcement learning systems for playing Atari games have had some impressive successes in transfer learning, but they still have not come close to learning to play new games as quickly as humans can.

    “Deep reinforcement learning systems for playing Atari games have had some impressive successes in transfer learning, but they still have not come close to learning to play new games as quickly as humans can.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation

    An attentional window is only a crude approximation of the true causal process of drawing with a pen, and in Rezende et al. (2016) it is not pen-like at all.

    “An attentional window is only a crude approximation of the true causal process of drawing with a pen, and in Rezende et al. (2016) the attentional window is not pen-like at all, although a more accurate pen model could be incorporated.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Rezende et al. 2016

      Not checked in the free run.

  • May combine two claims.

    Findingcitation

    Lopez-Paz et al.'s framework outperforms existing methods on various causal prediction tasks, but it is unclear how to apply it to inferring rich hierarchies of latent causal variables.

    “Although it outperforms existing methods on various causal prediction tasks, it is unclear how to apply the approach to inferring rich hierarchies of latent causal variables, as needed for the Frostbite Challenge and especially the Characters Challenge.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Lopez-Paz et al. 2015

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    People can quickly infer motor programs to draw a new character in a similarly guided process.

    “Similarly, as described in the Characters Challenge, people can quickly infer motor programs to draw a new character in a similarly guided processes.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingcitation

    The DRAW variant demonstrates an impressive ability to make plausible generalizations beyond training examples, yet it generalizes too broadly in other cases in non-human-like ways.

    “The model demonstrates an impressive ability to make plausible generalizations that go beyond the training examples, yet it generalizes too broadly in other cases, in ways that are not especially human-like.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Rezende et al. 2016

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given

    Many seemingly well-accepted ideas about neural computation are biologically dubious or at best uncertain.

    “Many seemingly well-accepted ideas regarding neural computation are in fact biologically dubious, or uncertain at best”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    Deep reinforcement learning is only just starting to address intrinsically motivated learning.

    “Deep reinforcement learning is only just starting to address intrinsically motivated learning (Kulkarni et al. 2016;Mohamed & Rezende 2015).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Kulkarni et al. 2016

      Not checked in the free run.

    • cites Mohamed & Rezende 2015

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    Neural networks are far from implementing human language abilities.

    “it is widely recognized that neural networks are far from implementing human language abilities.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    Exploration and creative innovation in the space of network architectures have not yet been made algorithmic.

    “Exploration and creative innovation in the space of network architectures have not yet been made algorithmic.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    Neuroscientists have not found mechanisms for implementing backpropagation in the brain, nor have they produced definitive evidence against it.

    “Although neuroscientists have not found any mechanisms for implementing backpropagation in the brain, neither have they produced definitive evidence against it.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Findingnone given

    Deep learning systems have not yet shown they can work with the structured representations discussed in the article, but they have demonstrated the surprising effectiveness of gradient descent in large models with high-dimensional parameter spaces.

    “Deep learning systems have not yet shown they can work with these representations, but they have demonstrated the surprising effectiveness of gradient descent in large models with high-dimensional parameter spaces.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation

    These abilities elude even the best deep neural networks for character recognition, which are trained on many examples of each concept and do not flexibly generalize to new tasks.

    “these abilities elude even the best deep neural networks for character recognition (Ciresan et al. 2012), which are trained on many examples of each concept and do not flexibly generalize to new tasks.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Ciresan et al. 2012

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    Current algorithms are still better at recognizing objects than understanding scenes, often getting the key objects right but their causal relationships wrong.

    “current algorithms are still better at recognizing objects than understanding scenes, often getting the key objects right but their causal relationships wrong (Fig. 6).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Fig. 6

      Not checked in the free run.

  • May combine two claims.

    Findingcitation

    Neural networks do build models, but these models are implicit and inseparable from the learning process.

    “neural networks do engage in model building, which is implicit, and cannot be dissociated from the learning process”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites Caglar & Hanson 2016

      Not checked in the free run.

    • cites Cleeremans 1993

      Not checked in the free run.

    • cites Hanson & Burr 1990

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    Lee Sedol has probably played around 50,000 games in his entire life.

    “Lee has probably played around 50,000 games in his entire life.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingcitation

    The DNC has been applied to solving block puzzles and finding paths between nodes in a graph after memorizing the graph.

    “the DNC has been applied to solving block puzzles and finding paths between nodes in a graph after memorizing the graph.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Graves et al. 2016

      Not checked in the free run.

  • May combine two claims.

    Findingreasoning

    People learn from fewer data and generalize in richer and more flexible ways than current best algorithms.

    “people learn from fewer data and generalize in richer and more flexible ways.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation

    AlphaGo is initially trained on 28.4 million positions and moves from 160,000 unique games played by human experts; it then improves through reinforcement learning, playing 30 million more games against itself.

    “AlphaGo is initially trained on 28.4 million positions and moves from 160,000 unique games played by human experts; it then improves through reinforcement learning, playing 30 million more games against itself.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Silver et al. 2016

      Not checked in the free run.

  • May combine two claims.

    Findingnone given

    AI can draw on observations of expert players but requires substantially greater input than humans to achieve similar performance.

    “AI can also draw on observations of expert players, but requires substantially greater input to achieve similar levels of performance.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation

    The supervisory system appears to perform functions such as goal generation, strategy generation and evaluation, monitoring, switching, and response inhibition.

    “The supervisory system appears to perform a variety of functions, such as goal generation (to create one's own goals and to function in real domains outside of the laboratory), strategy generation and evaluation (to create and evaluate potential strategies that might achieve goals), monitoring (to detect when one's goals are frustrated and to thereby trigger generation of new plans/strategies or”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites Shallice & Cooper [2011]

      Not checked in the free run.

    • cites Fox et al. [2013]

      Not checked in the free run.

    • cites Cooper [2016]

      Not checked in the free run.

  • May combine two claims.

    Findingcitation

    In the United States, China, and Iran, typically developing children improve in false-belief understanding over ontogeny, but the timing of improvement differs widely and may relate to different sociocultural inputs.

    “A large collection of research comparing the development of children's understanding of false belief in the United States, China, and Iran indicates that although typically developing children in all cultures show an improvement in false belief understanding over the course of ontogeny, the timing of this improvement differs widely-and such variability is potentially related to different socioc”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites Davoodi et al. 2016

      Not checked in the free run.

    • cites Liu et al. 2008

      Not checked in the free run.

    • cites Shahaeian et al. 2011

      Not checked in the free run.

  • May combine two claims.

    Findingnone given

    The differences between causal processes that generate a cat and those that generate a dog are understood imperfectly even by experts in developmental biology and hardly at all by laypeople.

    “Conversely, the differences between the causal processes that generate a cat and those that generate a dog are understood imperfectly, even by experts in developmental biology, and hardly at all by laypeople.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    Contemporary deep learning approaches to Atari games employ psychologically unrealistic training regimes and are inflexible to changes in reward/goal structure.

    “Although contemporary DL approaches to, say, learning and playing Atari games undoubtedly employ psychologically unrealistic training regimes, and are undoubtedly inflexible with respect to changes to the reward/ goal structure, to fixate on these limitations overlooks the promise of such approaches.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given

    Valence direction and magnitude often depend on the learner machine's disposition toward response priming/contagion, social facilitation, incentive motivation, and local/stimulus enhancement of observable demonstrator behavior.

    “Valence direction and magnitude often depend on the learner machine's disposition toward response priming/contagion, social facilitation, incentive motivation, and local/stimulus enhancement of observable demonstrator behavior”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingcitation

    Single-layer neural networks and multilayered deep learning networks have qualitatively different learning curve shapes: the former fits a negative exponential function (associative), while the latter fits a hyperbolic function (accumulative).

    “the shape of the learning curves for single-layer NNs and for multilayered DLs are qualitatively different, with the former fitting a negative exponential function (“associative”) and the latter fitting a hyperbolic function (“accumulative”)”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Caglar & Hanson 2016

      Not checked in the free run.

  • May combine two claims.

    Findingnone given

    There is no evidence that humans routinely generate detailed traces of positions, velocities, and forces over closely spaced instants, or that they are capable of doing so.

    “There is no evidence that humans routinely generate comparably detailed traces or even that they are capable of doing so.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingcitation

    Human subjects often make errors predicting the behavior of a balance beam, and most people find gyroscope behavior mystifying.

    “Human subjects often make errors in predicting the behavior of a balance beam (Siegler 1976), and most people find the behavior of a gyroscope mystifying.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Siegler 1976

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given

    Even children can readily distinguish dogs from cats.

    “Yet even children can readily distinguish dogs from cats.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Findingnone given

    Birds build species-typical nests they have never seen built, improving over time, and apes know a branch too weak to hold them.

    “Birds build species-typical nests they have never seen built, improving over time, and apes know a branch that is too weak to hold them.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingcitation

    Many successful models of intuitive physics have been built over decades, and these models have been used to ground scientific and engineering reasoning.

    “Decades of successful models have been built for many aspects of intuitive physics, and such models have also been used to ground scientific and engineering reasoning (Forbus 2011).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Forbus 2011

      Not checked in the free run.

  • May combine two claims.

    Findingcitation

    Companions can learn causal models by reading natural language texts, which improves their performance in a complex strategy game.

    “This has enabled Companions to learn causal models via reading natural language texts, thereby improving their performance in a complex strategy game (McFate et al. 2014).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites McFate et al. 2014

      Not checked in the free run.

  • May combine two claims.

    Findingcitation

    Studies suggest that people tend to distrust AIs and want to exert control and responsibility over the results such AIs deliver.

    “The studies by Dietvorst et al. (2015;2016;see Hutson 2017) suggest that people not only tend to distrust AIs, but also want to exert control, and hence responsibility, over the results such AIs deliver.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites Dietvorst et al. (2015

      Not checked in the free run.

    • cites [2016]

      Not checked in the free run.

    • cites Hutson 2017

      Not checked in the free run.

  • May combine two claims.

    Findingreasoning

    One's knowledge of another agent's mind and ability to predict their action are necessarily extremely limited, yet powerful psychological reasoning can be carried out.

    “One's knowledge of the state of another agent's mind and one's ability to predict their action are necessarily extremely limited; nonetheless, powerful psychological reasoning can be carried out.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation

    Physicists asked to solve a novel problem spontaneously use analogies to known systems, and frequent use of analogies is a major determinant of success in microbiology laboratories.

    “For example, physicists asked to solve a novel problem spontaneously use analogies to known systems (Clement 1988), and studies of working microbiology laboratories reveal that frequent use of analogies is a major determinant of success (Dunbar 1995).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Clement 1988

      Not checked in the free run.

    • cites Dunbar 1995

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    If you see a person pick up a telephone and dial, it is a good guess they are planning to talk to someone.

    “For example, if you see a person pick up a telephone and dial, it is a good guess that they he or she is planning to talk to someone.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingcitation

    Analogical processes support children's learning of physical science and mathematics.

    “Analogical processes support children's learning of physical science (Chen & Klahr 1999;Gentner et al. 2016) and mathematics (Carey 2009;Mix 1999;Richland & Simms 2015).”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites Chen & Klahr 1999

      Not checked in the free run.

    • cites Gentner et al. 2016

      Not checked in the free run.

    • cites Carey 2009

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    Simple changes to the visual environment, such as brightness, dramatically and adversely affect DeepQ network performance on breakout, whereas humans are not affected.

    “We found that even simple changes to the visual environment (as insignificant as changing the brightness) dramatically and adversely affect the performance of the algorithm, whereas humans are not affected by such perturbations at all.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    Model building-explaining observed data by constructing causal models of the world-is the hallmark of human-level learning.

    “we see model building as the hallmark of human-level learning, or explaining observed data through the construction of causal models of the world”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    People can recognize a new handwritten character from a single example and discriminate novel instances from similar non-instances.

    “People can learn to recognize a new handwritten character from a single example (Fig. 1A-i), allowing them to discriminate between novel instances drawn by other people and similar looking non-instances (Lake et al. 2015a;Miller et al. 2000).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Lake et al. 2015a

      Not checked in the free run.

    • cites Miller et al. 2000

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Equipped with general principles of intuitive physics, people learn more quickly and make more accurate predictions.

    “Equipped with these general principles, people can learn more quickly and make more accurate predictions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Model-free reinforcement learning algorithms, such as current deep RL, support fast control but at the cost of inflexibility and possibly accuracy.

    “By contrast, model-free reinforcement learning algorithms, such as current instantiations of deep reinforcement learning, support fast control, but at the cost of inflexibility and possibly accuracy.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    A primary job of learning is to extend and enrich these models and build analogous causally structured theories of other domains.

    “A primary job of learning is to extend and enrich these models and to build analogous causally structured theories of other domains.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Recent deep learning shares representational commitments and often learning algorithms with earlier PDP models.

    “The resurgence of recent interest in neural networks, more commonly referred to as”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites LeCun et al. [2015]

      Not checked in the free run.

    • cites Schmidhuber [2015]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Speech recognition is now dominated by fully neural network systems.

    “Now, the leading approaches to speech recognition are fully neural network systems (Graves et al. 2013;Hannun et al. 2014).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Graves et al. 2013

      Not checked in the free run.

    • cites Hannun et al. 2014

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    High-level feature representations of convolutional nets have been used to predict neural response patterns in human and macaque IT cortex, human typicality ratings, and similarity ratings for common object images.

    “where the high-level feature representations of these convolutional nets have also been used to predict patterns of neural response in human and macaque IT cortex (Khaligh-Razavi & Kriegeskorte 2014;Kriegeskorte 2015;Yamins et al. 2014), as well as human typicality ratings (Lake et al. 2015b) and similarity ratings (Peterson et al. 2016) for images of common objects.”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites Khaligh-Razavi & Kriegeskorte 2014

      Not checked in the free run.

    • cites Kriegeskorte 2015

      Not checked in the free run.

    • cites Yamins et al. 2014

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    This knowledge is often richly organized and theory-like, capable of graded inferences and productive capacities characteristic of human thought.

    “This knowledge is often richly organized and theory-like in structure, capable of the graded inferences and productive capacities characteristic of human thought.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The earlier a cognitive ingredient is present, the more likely it is foundational to later development and learning.

    “Also, the earlier an ingredient is present, the more likely it is to be foundational to later development and learning.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Compared with state-of-the-art machine learning algorithms, human learning is distinguished by its richness and efficiency.

    “Compared with state-of-the-art algorithms in machine learning, human learning is distinguished by its richness and its efficiency.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Children possess the ability and desire to uncover underlying causes of sparsely observed events and use that knowledge to go beyond limited data.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Planning in complex models is cumbersome and slow, making the speed-accuracy trade-off unfavorable for real-time control.

    “However, planning in complex models is cumbersome and slow, making the speedaccuracy trade-off unfavorable for real-time control.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Once a causal model of a task is learned, humans can use the model to plan action sequences that maximize future reward.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The claim that a mind is a collection of general-purpose neural networks with few initial constraints is rather extreme in contemporary cognitive science.

    “the claim that a mind is a collection of general-purpose neural networks with few initial constraints is rather extreme in contemporary cognitive science.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Deep learning models have yet to incorporate many of the essential ingredients for human-like learning and thinking, and may solve some problems differently than people.

    “finding that deep learning models have yet to incorporate many of them, and so may be solving some problems in different ways than people do.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: People are better than machines at solving a range of difficult computational problems including concept learning, scene understanding, language acquisition, language understanding, and speech recognition.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    Handwritten character recognition is simpler than more general object recognition because the object is two-dimensional, separated from background, and usually unoccluded.

    “Handwritten character recognition is also simpler than more general forms of object recognition; the object of interest is two-dimensional, separated from the background, and usually unoccluded.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Infants understand that other people have mental states such as goals and beliefs, and this understanding constrains their learning and predictions.

    “Infants understand that other people have mental states like goals and beliefs, and this understanding strongly constrains their learning and predictions.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Wellman & Gelman [1992]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Many algorithms achieve respectable performance on MNIST with large training data: KNN about 5% test error, SVM about 1%, and ConvNets below 1%.

    “With a large amount of training data available, many algorithms achieve respectable performance, including Knearest neighbors (5% test error), support vector machines (about 1% test error), and convolutional neural networks (below 1% test error [LeCun et al. 1998]).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites LeCun et al. 1998

      Not checked in the free run.

  • May combine two claims.

    Premisenone given

    A different picture has emerged in cognitive science highlighting early inductive biases, including core concepts such as number, space, agency, and objects, and powerful learning algorithms that rely on prior knowledge to extract knowledge from small amounts of training data.

    “A different picture has emerged that highlights the importance of early inductive biases, including core concepts such as number, space, agency, and objects, as well as powerful learning algorithms that rely on prior knowledge to extract knowledge from small amounts of training data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    If a cognitive ingredient is present early in development, it is active and available well before a child or adult would attempt to learn the tasks discussed.

    “If an ingredient is present early in development, it is certainly active and available well before a child or adult would attempt to learn the types of tasks discussed in this paper.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Some original PDP work was closer to model building than pattern recognition, while recent large-scale discriminative deep learning systems more purely exemplify pattern recognition.

    “Some of the original work done under the banner of PDP (Rumelhart et al. 1986b) is closer to model building than pattern recognition, whereas the recent large-scale discriminative deep learning systems more purely exemplify pattern recognition (see Bottou [2014] for a related discussion).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Rumelhart et al. 1986b

      Not checked in the free run.

    • cites Bottou [2014]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Convolutional networks have recently come close to human-level accuracy on certain object recognition benchmarks.

    “In the years since, ConvNets continue to dominate, recently approaching human-level performance on some object recognition benchmarks (He et al. 2016;Russakovsky et al. 2015;Szegedy et al. 2014).”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites He et al. 2016

      Not checked in the free run.

    • cites Russakovsky et al. 2015

      Not checked in the free run.

    • cites Szegedy et al. 2014

      Not checked in the free run.

  • May combine two claims.

    Interpretationcitation

    People learn from fewer examples and learn richer representations than neural networks, for both handwritten characters and more general object classes.

    “There are at least two important differences: people learn from fewer examples and they learn richer representations, a comparison true for both learning handwritten characters and for learning more general classes of objects (Fig. 1).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: People are better than machines at solving a range of difficult computational problems including concept learning, scene understanding, language acquisition, language understanding, and speech recognition.

  • May combine two claims.

    Premisecitation

    Infants know that objects persist over time and are solid and coherent.

    “For example, infants know that objects will persist over time and that they are solid and coherent.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Wellman & Gelman [1992]

      Not checked in the free run.

  • May combine two claims.

    Premisereasoning

    The DQN has very little built in apart from assumptions about image structure inherent in convolutional networks, so it must learn a visual and conceptual system from scratch for each new game.

    “The network embodies the strongly empiricist approach characteristic of most connectionist models: very little is built into the network apart from the assumptions about image structure inherent in convolutional networks, so the network has to essentially learn a visual and conceptual system from scratch for each new game.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Differences between human and machine learning curves suggest they may learn different kinds of knowledge, use different learning mechanisms, or both.

    “The differences between the human and machine learning curves suggest that they may be learning different kinds of knowledge, using different learning mechanisms, or both.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    Human learners fundamentally take on different learning tasks than today's neural networks.

    “Human learners fundamentally take on different learning tasks than today's neural networks”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    More recent DQN variants outperform the human tester, reaching 83% of the professional gamer's score with smarter experience replay and 172% with smarter replay and more efficient parameter sharing.

    “More recent variants of the DQN perform better, and can even outperform the human tester (Schaul et al. 2016;Stadie et al. 2016;van Hasselt et al. 2016;Wang et al. 2016), reaching 83% of the professional gamer's score by incorporating smarter experience replay (Schaul et al. 2016), and 172% by using smarter replay and more efficient parameter sharing (Wang et al. 2016)”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Schaul et al. 2016

      Not checked in the free run.

    • cites Wang et al. 2016

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    More recent work has shown game-specific networks can share visual features or train a multitask network, achieving modest transfer benefits when learning new games.

    “More recent work has shown how these game-specific networks can share visual features (Rusu et al. 2016) or be used to train a multitask network (Parisotto et al. 2016), achieving modest benefits of transfer when learning to play new games.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Rusu et al. 2016

      Not checked in the free run.

    • cites Parisotto et al. 2016

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    By the time a child or adult learns a new character or Frostbite, they have extensive real-world experience that deep learning systems do not benefit from.

    “By the time a child or an adult is picking up a new character or learning how to play Frostbite, he or she is armed with extensive real-world experience that deep learning systems do not benefit from”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    In addition to recognizing new examples, people can generate new examples, parse a character into important parts and relations, and generate new characters given a small set of related characters.

    “In addition to recognizing new examples, people can also generate new examples (Fig. 1A-ii), parse a character into its most important parts and relations (Fig. 1A-iii) (Lake et al. 2012), and generate new characters given a small set of related characters (Fig. 1A-iv).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Lake et al. 2012

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    People never start completely from scratch, or even close to it.

    “People never start completely from scratch, or even close to”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisenone given

    People can integrate across multiple examples of a character to infer optional elements, such as the cross-bar in '7's, and combine variants into a single coherent representation.

    “Furthermore, people can efficiently integrate across multiple examples of a character to infer which have optional elements, such as the horizontal cross-bar in '7's, combining different variants of the same character into a single coherent representation.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: People can integrate across multiple examples of a character to infer optional elements, such as the cross-bar in '7's, and combine variants into a single coherent representation.

  • May combine two claims.

    Premisecitation

    The DQN was a significant advance in reinforcement learning, showing a single algorithm can learn to play a wide variety of complex tasks.

    “The DQN was a significant advance in reinforcement learning, showing that a single algorithm can learn to play a wide variety of complex tasks.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Mnih et al. 2015

      Not checked in the free run.

  • May combine two claims.

    Interpretationnone given

    People understand enough to invent or accept new goals, generalize over input changes, and explain the game to others.

    “people understand enough to invent or accept new goals, generalize over changes to the input, and explain the game to others”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Recent experiments show that children seek out new data to distinguish hypotheses, isolate variables, test causal hypotheses, use the data-generating process, and learn selectively from others.

    “recent experiments showing that children seek out new data to distinguish between hypotheses, isolate variables, test causal hypotheses, make use of the data-generating process in drawing conclusions, and learn selectively from others”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites Cook et al. 2011

      Not checked in the free run.

    • cites Gweon et al. 2010

      Not checked in the free run.

    • cites Schulz et al. 2007

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    These core domains include number, space, physics, and psychology.

    “These domains include number (numerical and set operations), space (geometry and navigation), physics (inanimate objects and mechanics), and psychology (agents and groups)”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Although both people and models represent characters as sequences of pen strokes and relations, people have a richer repertoire of structural relations between strokes.

    “Although both people and models represent characters as a sequence of pen strokes and relations, people have a far richer repertoire of structural relations between strokes.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Important physical concepts appear at ages far earlier than when a child or adult learns Frostbite, suggesting these resources may be used for physics-related tasks.

    “important physical concepts are present at ages far earlier than when a child or adult learns to play Frostbite, suggesting these resources may be used for solving this and many everyday physics-related tasks”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationnone given

    People can learn models and use them for arbitrary new tasks and goals.

    “people can learn models and use them for arbitrary new tasks and goals”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    The core ingredients play an active and important role in producing human-like learning and thought, in ways contemporary machine learning has not captured.

    “the key claim is that these ingredients play an active and important role in producing human-like learning and thought, in ways contemporary machine learning has yet to capture”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The core ingredients of human intelligence are not present in most current learning-based AI systems, and not all of them are present together.

  • May combine two claims.

    Premisecitation

    The learning curve for Wang et al. (2016) shows performance of about 44% after 200 hours, 8% after 100 hours, and less than 2% after 5 hours, close to random play at about 1.5%.

    “The learning curve for the model of Wang et al. (2016) shows performance is approximately 44% after 200 hours, 8% after 100 hours, and less than 2% after 5 hours (which is close to random play, approximately 1.5%).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Wang et al. 2016

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    The DQN was trained on 200 million frames per game, equivalent to about 924 hours of game time, or almost 500 times the human's experience.

    “The DQN was trained on 200 million frames from each of the games, which equates to approximately 924 hours of game time (about 38 days), or almost 500 times as much experience as the human received.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Mnih et al. 2015

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Neural networks can learn multiple mappings or tasks from the same stimuli by adapting outputs to a specified goal, but adding new tasks requires substantial training or reconfiguration.

    “Although neural networks can learn multiple mappings or tasks with the same set of stimuliadapting their outputs depending on a specified goalthese models require substantial training or reconfiguration to add new tasks”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites Collins & Frank 2013

      Not checked in the free run.

    • cites Eliasmith et al. 2012

      Not checked in the free run.

    • cites Rougier et al. 2005

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    In Mnih et al. (2015), the network architecture and hyper-parameters were fixed, but the network was trained anew for each game, making the visual system and policy highly specialized for the games it was trained on.

    “In Mnih et al. (2015), the network architecture and hyper-parameters were fixed, but the network was trained anew for each game, meaning the visual system and the policy are highly specialized for the games it was trained on.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Mnih et al. 2015

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given

    The core domains are thought to be shared across cultures and partly with non-human animals.

    “together the domains are thought to be shared cross-culturally and partly with non-human animals”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationreasoning

    The core ingredients are enriched by experience, and some may even result from experience.

    “Certainly, the core ingredients are enriched by this experience, and some may even be a product of the experience itself”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Liquids are expected to pass through barriers, while solid objects cannot.

    “Liquids, for example, are expected to go through barriers, while solid objects cannot”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Hespos et al. 2009

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Young infants expect objects to move along smooth paths, not vanish and reappear, not interpenetrate, and not act at a distance.

    “Young infants believe objects should move along smooth paths, not wink in and out of existence, not inter-penetrate and not act at a distance”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisenone given

    Intuitive psychology abilities are early emerging and play an important role in human learning and thought.

    “it is clear that these abilities are early emerging and play an important role in human learning and thought”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisenone given

    Intuitive psychology is an early-emerging ability that importantly influences human learning and thought.

    “Intuitive psychology is another early-emerging ability with an important influence on human learning and thought”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given

    An explicit agent representation can predict how a sidekick agent will be helpful in new circumstances, whereas a bottom-up pixel-based representation is likely to struggle.

    “An explicit agent representation can predict how such an agent will be helpful in new circumstances, whereas a bottom-up pixel-based representation is likely to struggle”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    People require far less experience than PhysNet to perform any particular task and can generalize to many novel judgments and complex scenes without new training.

    “In contrast, people require far less experience to perform any particular task, and can generalize to many novel judgments and complex scenes with no new training required”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    Any full formal account of intuitive psychological reasoning needs representations of agency, goals, efficiency, and reciprocal relations.

    “it seems to us that any full formal account of intuitive psychological reasoning needs to include representations of agency, goals, efficiency, and reciprocal relations”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    Simulation-based reasoning in intuitive psychology can be nested recursively to understand social interactions, unlike in intuitive physics.

    “Importantly, unlike in intuitive physics, simulation-based reasoning in intuitive psychology can be nested recursively to understand social interactions”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisenone given

    Such cues are often sufficient but not necessary for detecting agency.

    “Such cues are often sufficient but not necessary for the detection of agency”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisenone given

    These connectionist networks do not attempt to work with complex scenes as input or a wide range of scenarios and judgments.

    “However, these networks do not attempt to work with complex scenes as input, or a wide range of scenarios and judgments as in Figure 4”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The DQN probably does not parse a Frostbite screenshot in terms of stable objects or sprites moving according to intuitive physics.

    “the DQN probably does not parse a Frostbite screenshot in terms of stable objects or sprites moving according to the rules of intuitive physics”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    According to the intuitive physics engine hypothesis, people reconstruct a perceptual scene using internal representations of objects, their physical properties, and forces acting on them.

    “According to this hypothesis, people reconstruct a perceptual scene using internal representations of the objects and their physically relevant properties (such as mass, elasticity, and surface friction) and forces acting on objects (such as gravity, friction, or collision impulses)”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing), S4 (based).

    Interpretationnone given

    The intuitive physics engine approach enables flexible adaptation to many everyday scenarios and judgments beyond perceptual cues.

    “This”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    By simulating planning processes, people can predict what agents might do next or infer utilities and beliefs from observed actions.

    “By simulating these planning processes, people can predict what agents might do next, or use inverse reasoning from observing a series of actions to infer the utilities and beliefs of agents in a scene”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    Neural network learning algorithms implement learning as gradual adjustment of connection strengths.

    “these algorithms implement learning as a process of gradual adjustment of connection strengths”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    For supervised learning, updates usually aim to improve pattern recognition capabilities.

    “For supervised learning, the updates are usually aimed at improving the algorithm's pattern recognition capabilities”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    These models formalize mentalistic concepts such as goal, agent, planning, cost, efficiency, and belief.

    “These models formalize explicitly mentalistic concepts such as”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisenone given

    The intuitive physical state representation is approximate, probabilistic, oversimplified, and incomplete relative to physical ground truth.

    “Relative to physical ground truth, the intuitive physical state representation is approximate and probabilistic, and oversimplified and incomplete in many ways”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Infants expect agents to act contingently and reciprocally, to have goals, and to take efficient actions toward those goals subject to constraints.

    “infants also expect agents to act contingently and reciprocally, to have goals, and to take efficient actions toward those goals subject to constraints”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites Csibra 2008

      Not checked in the free run.

    • cites Csibra et al. 2003

      Not checked in the free run.

    • cites Spelke & Kinzler 2007

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Relating data to their causal source provides strong priors for perception and learning, and a richer basis for generalizing in new ways and to new tasks.

    “Relating data to their causal source provides strong priors for perception and learning, as well as a richer basis for generalizing in new ways and to new tasks.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    There are domains where machine learners outperform human learners, such as combing through financial or weather data.

    “There are also domains for which machine learners outperform human learners, such as combing through financial or weather data”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Concepts support prediction, action, communication, imagination, explanation, and composition beyond classification.

    “Beyond classification, concepts support prediction (Murphy & Ross 1994;Rips 1975), action (Barsalou 1983), communication (Markman & Makin 1998), imagination (Jern & Kemp 2013;Ward 1994), explanation (Lombrozo 2009;Williams & Lombrozo 2010), and composition (Murphy 1988;Osherson & Smith 1981)”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites Murphy & Ross 1994

      Not checked in the free run.

    • cites Barsalou 1983

      Not checked in the free run.

    • cites Markman & Makin 1998

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Networks trained for object recognition encode part-like features in their deeper layers, and new objects can activate novel combinations of feature detectors.

    “Networks trained for object recognition encode part-like features in their deeper layers (Zeiler & Fergus 2014), whereby the presentation of new types of objects can activate novel combinations of feature detectors.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Zeiler & Fergus 2014

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Handwritten characters are inherently compositional, with parts being pen strokes and relations describing how strokes connect.

    “Handwritten characters are inherently compositional, where the parts are pen strokes, and relations describe how these strokes connect to each other.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    MNIST includes 6,000 examples of each handwritten digit.

    “the MNIST benchmark includes 6,000 examples of each handwritten digit”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites LeCun et al. 1998

      Not checked in the free run.

  • May combine two claims.

    Interpretationnone given

    At least for image and caption generation, recent tasks have mostly been studied in the big data setting, which conflicts with the human ability to generalize from small data sets.

    “Yet, at least for image and caption generation, these tasks have been mostly studied in the big data setting that is at odds with the impressive human ability to generalize from small data sets”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    These abilities are not independent; they interact and come for free with acquisition of the underlying concept.

    “These abilities are not independent; rather they hang together and interact (Solomon et al. 1999), coming for free with the acquisition of the underlying concept”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Solomon et al. 1999

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    In both the MNIST and Atari DQN cases, algorithms use information less efficiently than a person learning the same tasks.

    “In both cases, the algorithms are clearly using information less efficiently than a person learning to perform the same tasks”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    A scene from the game Frostbite is a composition of various object types, including birds, fish, ice floes, and igloos.

    “A scene from the game is a composition of various object types, including birds, fish, ice floes, igloos, and so on (Fig. 2).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    These BPL programs allow the model to express causal knowledge about how raw data are formed, and their probabilistic semantics allow handling noise and creative tasks.

    “These programs allow the model to express causal knowledge about how the raw data are formed, and the probabilistic semantics allow the model to handle noise and perform creative tasks”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Representing compositional structure explicitly is more economical and better for generalization.

    “Representing this compositional structure explicitly is both more economical and better for generalization, as noted in previous work on object-oriented reinforcement learning (Diuk et al. 2008).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Diuk et al. 2008

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Compositionality, causality, and learning-to-learn were important to the success of the BPL framework and are important to understanding human learning as rapid model building more broadly.

    “The following sections discuss the three main ingredientscompositionality, causality, and learning-to-learnthat were important to the success of this framework and, we believe, are important to understanding human learning as rapid model building more broadly.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The success of the BPL framework in the Characters Challenge can be generalized to understanding human learning as rapid model building.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Lopez-Paz et al. introduced a discriminative, data-driven framework for distinguishing the direction of causality from examples.

    “Lopez-Paz et al. (2015) introduced a discriminative, data-driven framework for distinguishing the direction of causality from examples.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Lopez-Paz et al. 2015

      Not checked in the free run.

  • May combine two claims.

    Premisereasoning

    Concept learning and vision models that use causality are usually generative, but not every generative model is causal.

    “Concept learning and vision models that use causality are usually generative (as opposed to discriminative; see Glossary in Table 1), but not every generative model is also causal.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Liberman et al. argued that the richness of speech perception is best explained by inverting the production plan at the level of vocal tract movements.

    “For example, Liberman et al. (1967) argued that the richness of speech perception is best explained by inverting the production plan, at the level of vocal tract movements, to explain the large amounts of acoustic variability and the blending of cues across adjacent phonemes.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Liberman et al. 1967

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given

    Deep learning has efficient inference and scalable learning as important strengths.

    “It also suggests an additional way to build on the successes of deep learning, where efficient inference and scalable learning are important strengths of the approach.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Hierarchical Bayesian models operating over probabilistic programs are equipped to deal with theory-like structures and rich causal representations of the world, yet there are formidable algorithmic challenges for efficient inference.

    “Hierarhical Bayesian models operating over probabilistic programs (Goodman et al. 2008;Lake et al. 2015a;Tenenbaum et al. 2011) are equipped to deal with theory-like structures and rich causal representations of the world, yet there are formidable algorithmic challenges for efficient inference.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites Goodman et al. 2008

      Not checked in the free run.

    • cites Lake et al. 2015a

      Not checked in the free run.

    • cites Tenenbaum et al. 2011

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisereasoning

    People assume that interacting with birds and fish will change the game state in some way, because video games typically yield costs or rewards for these types of interactions.

    “Whether the person is a child or a seasoned gamer, it seems obvious that interacting with the birds and fish will change the game state in some way, either good or bad, because video games typically yield costs or rewards for these types of interactions (e.g., dying or points).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    A causal model of Frostbite would have to be more complex, gluing together object representations and explaining their interactions with intuitive physics and intuitive psychology.

    “A causal model of Frostbite would have to be more complex, gluing together object representations and explaining their interactions with intuitive physics and intuitive psychology, much like the game engine that generates the game dynamics and, ultimately, the frames of pixel images.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    If such a system aims to learn compositionally structured causal models of each game, built on a foundation of intuitive physics and psychology, it could transfer knowledge more efficiently and learn new games much more quickly.

    “Yet, if such a system aims to learn compositionally structured causal models of each gamebuilt on a foundation of intuitive physics and psychologyit could transfer knowledge more efficiently and thereby learn new games much more quickly.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given

    Representing intuitive theories and structured causal models is less natural in deep neural networks, but recent progress has demonstrated the remarkable effectiveness of gradient-based learning in high-dimensional parameter spaces.

    “In contrast, whereas representing intuitive theories and structured causal models is less natural in deep neural networks, recent progress has demonstrated the remarkable effectiveness of gradient-based learning in high-dimensional parameter spaces.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Incorporating causality may greatly improve deep learning models, which were trained without access to causal data about how characters are produced.

    “Incorporating causality may greatly improve these deep learning models; they were trained without access to causal data about how characters are actually produced, and without any incentive to learn the true causal process.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    In general, richer and more structured models require more complex and slower inference algorithms.

    “In general, richer and more structured models require more complex and slower inference algorithms, similar to how complex models require more data, making the speed of perception and thought all the more remarkable.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisenone given

    Computing a probability distribution over an entire space of programs is usually intractable, and often even finding a single high-probability program poses an intractable search problem.

    “Computing a probability distribution over an entire space of programs is usually intractable, and often even finding a single high-probability program poses an intractable search problem.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    People immediately parse the game environment into objects, types of objects, and causal relations between them.

    “Most basically, they immediately parse the game environment into objects, types of objects, and causal relations between them.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning

    BPL could benefit from deeper forms of learning-to-learn than it currently does.

    “BPL could also benefit from deeper forms of learning-tolearn than it currently does.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Graves (2014) learned a generative model of cursive handwriting using a recurrent neural network trained on handwriting data.

    “Graves ( 2014) learned a generative model of cursive handwriting using a recurrent neural network trained on handwriting data.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Graves 2014

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    Capturing more human-like learning-to-learn dynamics in deep networks and other machine learning approaches could facilitate much stronger transfer to new tasks and new problems.

    “Capturing more human-like learning-to-learn dynamics in deep networks and other machine learning approaches could facilitate much stronger transfer to new tasks and new problems.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning

    Some of the important structure BPL exploits to generalize well is built into the prior and not learned from background pre-training, whereas people might learn this knowledge.

    “Some of the important structure it exploits to generalize well is built in to the prior and not learned from the background pre-training, whereas people might learn this knowledge, and ultimately, a human-like machine learning system should as well.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    It has been proposed that humans can approximate Bayesian inference using Monte Carlo methods.

    “Most prominently, it has been proposed that humans can approximate Bayesian inference using Monte Carlo methods, which stochastically sample the space of possible hypotheses and evaluate these samples according to their consistency with the data and prior knowledge (Bonawitz et al. 2014;Gershman et al. 2012;Ullman et al. 2012b;Vul et al. 2014).”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites Bonawitz et al. 2014

      Not checked in the free run.

    • cites Gershman et al. 2012

      Not checked in the free run.

    • cites Ullman et al. 2012b

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    Deep neural networks could play a role in two ways: as a bottom-up proposer to make probabilistic inference more tractable in a structured generative model, or as the causal generative model if imbued with the right ingredients.

    “Deep neural networks could play a role in two ways: by serving as a bottom-up proposer to make probabilistic inference more tractable in a structured generative model (sect.4.3.1)or by serving as the causal generative model if imbued with the right set of ingredients.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    People understand that video games like Frostbite have goals, which often involve approaching or avoiding objects based on their type.

    “People also understand that video games like these have goals, which often involve approaching or avoiding objects based on their type.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    If deep neural networks could adopt similarly compositional, hierarchical, and causal representations, they could benefit more from learning-to-learn.

    “If deep neural networks could adopt similarly compositional, hierarchical, and causal representations, we expect they could benefit more from learningto-learn.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    In hierarchical Bayesian modeling, a general prior on concepts is shared by multiple specific concepts, and the prior itself is learned over the course of learning the specific concepts.

    “In hierarchical Bayesian modeling (Gelman et al. 2004), a general prior on concepts is shared by multiple specific concepts, and the prior itself is learned over the course of learning the specific concepts (Salakhutdinov et al. 2012;2013).”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites Gelman et al. 2004

      Not checked in the free run.

    • cites Salakhutdinov et al. 2012

      Not checked in the free run.

    • cites Salakhutdinov et al. 2013

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    The strong priors, constraints, or inductive biases needed to learn a particular task quickly are often shared to some extent with other related tasks.

    “The strong priors, constraints, or inductive bias needed to learn a particular task quickly are often shared to some extent with other related tasks.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationcitation

    This amortization process might occur offline, suggesting a form of consolidation in reinforcement learning.

    “This process might occur offline (e.g., in dreaming or quiet wakefulness), suggesting a form of consolidation in reinforcement learning (Gershman et al. 2014).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Gershman et al. 2014

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Some evidence for inferential correlations in humans was reported by Gershman and Goodman (2014).

    “Some evidence for inferential correlations in humans was reported by Gershman and Goodman (2014).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Gershman and Goodman 2014

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationcitation

    Another avenue for potential integration is through differentiable programming, by ensuring that program-like hypotheses are differentiable and thus learnable via gradient descent.

    “Another avenue for potential integration is through differentiable programming (Dalrymple 2016), by ensuring that the programlike hypotheses are differentiable and thus learnable via gradient descenta possibility discussed in the concluding section (Section 6.1).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Dalrymple 2016

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Language facilitates more powerful learning-to-learn and compositionality, allowing people to learn more quickly and flexibly by representing new concepts and thoughts in relation to existing concepts.

    “Language also facilitates more powerful learning-to-learn and compositionality (Mikolov et al. 2016), allowing people to learn more quickly and flexibly by representing new concepts and thoughts in relation to existing concepts”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites Mikolov et al. 2016

      Not checked in the free run.

    • cites Lupyan & Bergen 2016

      Not checked in the free run.

    • cites Lupyan & Clark 2015

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Understanding language and its role in intelligence goes hand-in-hand with understanding the building blocks discussed in the article.

    “understanding language and its role in intelligence goes hand-in-hand with understanding the building blocks discussed in this article.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    A competent Frostbite player can easily shift behavior appropriately, with little or no additional learning, implying a model-based planning approach.

    “We conjecture that a competent Frostbite player can easily shift behavior appropriately, with little or no additional learning, and it is hard to imagine a way of doing that other than having a model-based planning approach in which the environment model can be modularly combined with arbitrary new reward functions and then deployed immediately for planning.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    A marriage of flexibility and efficiency might be achievable if we use the human reinforcement learning systems as guidance.

    “Thus, a marriage of flexibility and efficiency might be achievable if we use the human reinforcement learning systems as guidance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Behavior can persist for weeks or months, whereas LTP decays to baseline over the course of days.

    “Behavior can persist for weeks or months, whereas LTP decays to baseline over the course of days”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Power et al. 1997

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Deep learning models could incorporate the core cognitive ingredients through additional structure and perhaps additional learning mechanisms, but most have not yet done so.

    “Deep learning models could incorporate these ingredients through some combination of additional structure and perhaps additional learning mechanisms, but for the most part have yet to do so.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Schulz (2012b) has suggested that abstract structural properties of problems contain information about the abstract forms of their solutions.

    “Schulz (2012b) has suggested that abstract structural properties of problems contain information about the abstract forms of their solutions.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Schulz 2012b

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Learned behavior is rapidly re-acquired after extinction, whereas no such facilitation is observed for LTP.

    “Learned behavior is rapidly re-acquired after extinction (Bouton 2004), whereas no such facilitation is observed for LTP (Jonge & Racine 1985).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Bouton 2004

      Not checked in the free run.

    • cites Jonge & Racine 1985

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    There may also be an intrinsic drive to reduce uncertainty and construct models of the environment.

    “There may also be an intrinsic drive to reduce uncertainty and construct models of the environment (Edelman 2015;Schmidhuber 2015), closely related to learning-to-learn and multitask learning.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Edelman 2015

      Not checked in the free run.

    • cites Schmidhuber 2015

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning

    Any approach to human-like AI, whether based on deep learning or not, is likely to benefit from incorporating the core cognitive ingredients.

    “Any approach to human-like AI, whether based on deep learning or not, is likely to gain from incorporating these ingredients.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    These capacities are in place before children master language, and they provide the building blocks for linguistic meaning and language acquisition.

    “These capacities are in place before children master language, and they provide the building blocks for linguistic meaning and language acquisition”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites Carey 2009

      Not checked in the free run.

    • cites Jackendoff 2003

      Not checked in the free run.

    • cites Kemp 2007

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    Autonomous cars could more accurately predict pedestrian behavior by inferring mental states, including their beliefs and desires.

    “autonomous cars could more accurately predict pedestrian behavior by inferring mental states, including their beliefs (e.g., Do they think it is safe to cross the street?Are they paying attention?)and desires”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    Like AlphaGo, people are also integrating model-free pattern recognition with model-based search.

    “And like AlphaGo, they are also integrating model-free pattern recognition with model-based search.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    Causality helps infuse the recognition of existing tools or the learning of new ones with an understanding of their use, helping to connect different object models in the proper way.

    “causality helps infuse the recognition of existing tools or the learning of new ones with an understanding of their use, helping to connect different object models in the proper way”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    A human-like system would be able to learn both to recognize and to produce new instances from a small number of examples.

    “a human-like system would be able to learn both to recognize and to produce new instances from a small number of examples.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    Studying the brain can provide valuable insights for tackling the architecture challenge.

    “looking at the brain can furnish important insights on how to face the architecture challenge”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Robots and personal assistants cannot be pretrained on all possible concepts they may encounter.

    “Robots and personal assistants such as cell phones cannot be pretrained on all possible concepts they may encounter.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The human brain resembles a complex society of diverse neural networks more than a single large network.

    “the human brain is more similar to a complex society of diverse neural networks, than to a large, single one”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Minsky 1986

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Deep learning and other learning paradigms can move closer to human-like learning and thought if they incorporate psychological ingredients.

    “deep learning and other learning paradigms can move closer to human-like learning and thought if they incorporate psychological ingredients”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning

    This type of psychological reasoning, along with other types of model-based causal and physical reasoning, are likely to be especially valuable in challenging and novel driving circumstances for which there are few relevant training data.

    “This type of psychological reasoning, along with other types of model-based causal and physical reasoning, are likely to be especially valuable in challenging and novel driving circumstances for which there are few relevant training data”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationreasoning

    A system may be able to quickly learn new concepts by constructing them from pre-existing primitive actions, informed by knowledge of the underlying causal process and learning-to-learn.

    “a system may be able to quickly learn new concepts by constructing them from pre-existing primitive actions, informed by knowledge of the underlying causal process and learning-to-learn.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given

    People are using their core intuitive psychology and aspects of their intuitive physics.

    “They are using their core intuitive psychology and aspects of their intuitive physics (spatial and object representations).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    AlphaGo cannot even conceive of these variants, let alone adapt to them autonomously, which is a sign that it does not understand the game as humans do.

    “the fact that it cannot even conceive of these variants, let alone adapt to them autonomously, is a sign that it does not understand the game as humans do.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    People are learning-to-learn with compositional knowledge.

    “They are learning-tolearn with compositional knowledge.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Compositionality, causality, intuitive physics, and intuitive psychology play an increasingly important role in reaching true scene understanding.

    “We see compositionality, causality, intuitive physics, and intuitive psychology as playing an increasingly important role in reaching true scene understanding.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    The two brain hemispheres appear to use two opposing strategies: an analogical-intuitive, gradient-descent-like one, and a digital-rational, vector-quantization-like one.

    “The two hemispheres seem to follow two opposite criteria: an analogicalintuitive one, gradient descent-like, and a digital-rational one, vector quantization-like.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning

    Perfect autonomous driving requires intuitive psychology.

    “Perfect autonomous driving requires intuitive psychology.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    In learning to play Go as quickly and as flexibly as they do, people are drawing on most of the cognitive ingredients this article has laid out.

    “In learning to play Go as quickly and as flexibly as they do, people are drawing on most of the cognitive ingredients this article has laid out.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    Other drivers on the road have similarly complex mental states underlying their behavior.

    “other drivers on the road have similarly complex mental states underlying their behavior”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Human intelligence operates as a prediction engine that generates hypotheses from a relatively small database and continuously verifies them through sequential sampling.

    “Human intelligence is a prediction engine working on hypotheses, generated from a relatively small database and constantly verified through sequential sampling”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    A large portion of the brain's intelligence is located in its architecture.

    “we think that much of the intelligence of the brain resides in its architecture”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    If a scene includes people acting or interacting, it will be nearly impossible to understand their actions without thinking about their thoughts and especially their goals and intentions toward the other objects and agents they believe are present.

    “If the scene includes people acting or interacting, it will be nearly impossible to understand their actions without thinking about their thoughts and especially their goals and intentions toward the other objects and agents they believe are present.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    For an autonomous agent to effectively navigate and perform tasks in a cluttered environment, the agent would need intuitive physics to properly reason about stability and support.

    “For an autonomous agent to effectively navigate and perform tasks in this environment, the agent would need intuitive physics to properly reason about stability and support.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationreasoning

    A holistic model of a scene would require the composition of individual object models, glued together by relations.

    “A holistic model of the scene would require the composition of individual object models, glued together by relations.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    In creative design, the free combination of parts is not enough on its own: compositionality and learning-to-learn can provide the parts for new ideas, but causality provides the glue that gives them coherence and purpose.

    “In each case, the free combination of parts is not enough on its own: Although compositionality and learning-to-learn can provide the parts for new ideas, causality provides the glue that gives them coherence and purpose.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    Natural intelligence is still by far the best example of intelligence.

    “natural intelligence is still by far the best example of intelligence.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Premisenone given

    Human-like learning and decision making rely on rich internal models.

    “Human-like learning and decision making surely do depend upon rich internal models”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning

    It is uncertain whether detailed knowledge engineering can be realistically achieved in all domains we want agents to handle.

    “it is not clear that detailed knowledge engineering will be realistically attainable in all areas we will want our agents to tackle”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given

    Machine performance may rival or exceed human performance on particular tasks, but it does not follow that the algorithm learns or thinks like a person.

    “Machine performance may rival or exceed human performance on particular tasks, and algorithms may take inspiration from neuroscience or aspects of psychology, but it does not follow that the algorithm learns or thinks like a person.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationreasoning

    Future AI systems will need sophisticated architectures, and knowledge of the brain could help guide their construction.

    “Future artificialintelligence systems will require sophisticated architectures, and knowledge of the brain might guide their construction”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Premisenone given

    Prior knowledge must inform and constrain the learning process.

    “the learning process must be informed and constrained by prior knowledge”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Knowledge is based on exemplar-based learning and generalization, combined with highly flexible generalization.

    “knowledge is grounded in exemplar-based learning and generalization, combined with high flexible generalization”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Natural intelligence appears to combine the largely reactive perception-action cycle of RL with a more active action-perception cycle in which the cognitive system can act and deliberatively explore to test hypotheses.

    “Natural intelligence appears to combine the largely reactive perception-action cycle of RL with a more active action-perception cycle, in which the cognitive system can act and deliberatively explore in order to test hypotheses.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    In intuitive psychology, many inferences proceed without detailed causal generative models.

    “In intuitive psychology, many inferences proceed without detailed causal generative models.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    Lake et al.'s results may tell us little about object recognition in general.

    “Lake et al.'s results are technically impressive, but may tell us little about object recognition in general.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The promise of DL networks is that with sufficiently broad training they may support systems that capture intuitive physics and intuitive psychology.

    “The promise of DL networks (and at present it remains a promise) is that, with sufficiently broad training, they may support the development of systems that capture intuitive physics and intuitive psychology.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    In physical reasoning, a model can be a set of constraints rather than a physics engine.

    “In physical reasoning, a model can be a set of constraints rather than a physics engine.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    Over time, these culturally dependent updates may result in distinct operating systems.

    “Over time, these updates may even result in distinct operating systems.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationcitation

    The routine system may plausibly employ a DL-type network combined with model-free reinforcement learning, while the non-routine system is more plausibly conceived as a model-based system.

    “Within this account the routine system may plausibly employ a DL-type network combined with (a hierarchical variant of) model-free reinforcement learning, whereas the non-routine system is more plausibly conceived of in terms of a model-based system (cf.Daw et al. 2005).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Daw et al. 2005

      Not checked in the free run.

  • May combine two claims.

    Premisenone given

    Simulation engines require both a precise predictive theory of the domain and a geometrically and physically precise description of the situation.

    “First, simulation engines require both a precise predictive theory of the domain and a geometrically and physically precise description of the situation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Accounts of behavior generation and regulation have long distinguished between automatic and deliberative behavior.

    “More specifically, accounts of behaviour generation and regulation have long distinguished between automatic and deliberative behaviour.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    The contention scheduling/supervisory system theory proposes one system controls routine, overlearned, or automatic behavior, while a second system may bias or modulate it in non-routine situations.

    “Thus, the contention scheduling/supervisory system theory of Norman and Shallice (1986) proposes that one systemthe contention scheduling systemcontrols routine, overlearned, or automatic behaviour, whereas a second systemthe supervisory systemmay bias or modulate the contention scheduling system in non-routine situations where deliberative control is exercised.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Norman and Shallice (1986)

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Children's developmental start-up software is shaped by their culture-specific social environment.

    “What is missing from this discussion, however, is the recognition that children's developmental start-up software is shaped by their culture-specific social environment.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    The world knowledge used in vision presumably involves a combination of geometric, physical, and other knowledge, rather than just a causal theory of how the image was produced.

    “For example, the world knowledge used in vision presumably involves a combination of geometric, physical, and other knowledge, rather than just a causal theory of how the image was produced.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Neither the balance beam errors nor gyroscope mystification follows from the physics engine model.

    “Neither result follows from the model.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Analogical comparison processes are central to human cognition.

    “We argue that (1) analogical comparison processes are central to human cognition; and (2) intuitive physical knowledge is captured by qualitative representations, rather than quantitative simulations.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Premisereasoning

    AGI has one existence proof: our brains.

    “AGI has one existence proofour brains.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisenone given

    Qualitative representations give symbolic relational representations of continuous properties and an account of causality organized around processes of change.

    “Qualitative representations provide symbolic, relational representations of continuous properties and an account of causality organized around processes of change.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Premisenone given

    Analogy is a way to transfer knowledge from prior situations.

    “Analogy provides a means of transferring prior knowledge.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationcitation

    Even visual tasks such as character recognition are more compactly represented by a network of relationships and objects than by an array of pixels, which is why human visual systems compute edges.

    “Even visual tasks, such as character recognition, are more compactly represented by a network of relationships and objects than by an array of pixels, which is why human visual systems compute edges (Marr 1983;Palmer 1999).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Marr 1983

      Not checked in the free run.

    • cites Palmer 1999

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationcitation

    An AI could not be created to have genuine interests, but that is down a very long road.

    “We are not saying that an AI could not be created to have genuine interests, but that is down a very long road (Dennett 2017;Hurley et al. 2011).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Dennett 2017

      Not checked in the free run.

    • cites Hurley et al. 2011

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    Long before AIs with genuine interests arrive, if ever, we will have AIs that can communicate with natural language with their users.

    “long before that day arrives, if it ever does, we will have AIs that can communicate with natural language with their users (not collaborators).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Some promising current work suggests that genuine human consciousness depends on a fundamental architecture that would require having interests.

    “Although some promising current work suggests that genuine human consciousness depends on a fundamental architecture that would require having interests (Deacon 2012;Dennett 2013), long before that day arrives, if it ever does, we will have AIs that can communicate with natural language with their users (not collaborators).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Deacon 2012

      Not checked in the free run.

    • cites Dennett 2013

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Because structure mapping allows relational representations, the system also provided explanations, which distributed representations lack.

    “Because structure mapping allows the use of relational representations, the system also provided explanations, the lack of which is a significant drawback of distributed representations.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisenone given

    Spatial lateral connections between oriented features are a predominant feature of visual cortex and are known to help enforce contour continuity.

    “spatial lateral connections between oriented features are a predominant feature of the visual cortex and are known to play a role in enforcing contour continuity.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Analogy can both explain human transfer learning and offer new techniques for machine learning.

    “This and many other experiments suggest that analogy not only can explain human transfer learning, but also can provide new techniques for machine learning.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationcitation

    Qualitative models can explain aspects of social reasoning, including blame assignment and moral decision making, suggesting they matter for intuitive psychology.

    “Moreover, qualitative models can explain aspects of social reasoning, including blame assignment (Tomai & Forbus 2008) and moral decision making (Dehghani et al. 2008), suggesting that they are important in intuitive psychology as well.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Tomai & Forbus 2008

      Not checked in the free run.

    • cites Dehghani et al. 2008

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    Human cognition relies heavily on structured relational representations.

    “We also agree that human cognition relies heavily on structured relational representations.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning

    Even the most powerful existing AIs are intelligent tools, not colleagues, and should not be granted moral authority or responsibility because they do not have interests, and simulated interests are not enough.

    “Even the most powerful of existing AIs are intelligent tools, not colleagues, and whereas they can be epistemically authoritative (within limits we need to characterize carefully), and hence will come to be relied on more and more, they should not be granted moral authority or responsibility because they do not have skin in the game: they do not yet have interests, and simulated interests are not”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Rapid computation over structured representations is a major feature of human cognition.

    “we agree that rapid computation over structured representations is a major feature of human cognition.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    Humans eventually bootstrap behavioral competence into reflective comprehension, something more like a theory and probably beyond language-less animals.

    “We humans eventually bootstrap this behavioral competence into reflective comprehension, something more like a theory and something that is probably beyond language-less animals.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    It is difficult to see how a Monte Carlo simulation approach would capture either the semantics of language about processes or the findings of the conceptual change literature.

    “It is hard to see how a Monte Carlo simulation approach would capture either the semantics of language about processes or the findings of the conceptual change literature.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    Human emotion is probably the biggest difference between people and AI machines.

    “Human emotion is probably the biggest difference between people and AI machines.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    Not reaching an important goal can make a person angry.

    “For example, not reaching an important goal can make a person angry.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisereasoning

    Emotions modulate human behavior, and the how of behavior is determined by emotions.

    “Emotions modulate human behavior; the how of the behavior is determined by the emotions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    This routing protocol should be relatively fixed in a given species and lineage and scalable over development and evolution.

    “This protocol should be relatively fixed in a given species and lineage, and have the ability to be scaled up over development and evolution.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisereasoning

    The brain appears to largely show dynamic routing, sparse connectivity, and feedback among many hierarchical levels.

    “the brain appears to largely show dynamic routing, sparse connectivity, and feedback among many hierarchical levels.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Human intelligence controls itself by reprogramming its own programs.

    “In this sense, human intelligence controls itself, reprogramming its own programs.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    MCMC has asymptotic guarantees.

    “Although MCMC has asymptotic guarantees”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationnone given

    Motivation can explain why a plan has priority, why it is executed, or why an action is stopped.

    “Motivation can explain why a certain plan has priority and why it is executed, or why a certain action is stopped.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    Cortical neurons, like Internet hosts, are separated by short network distances and can communicate along many possible routes within milliseconds.

    “Cortical neurons, like Internet hosts, are separated by relatively short network distances, and have the potential for communication along many possible routes within milliseconds.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning

    Brains are expected to adopt analogous, if not homologous, solutions when conditions require.

    “We should expect that brains adopt analogousif not homologoussolutions when conditions require.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Including dynamic routing, sparse connectivity, and hierarchical feedback in computational models may better approximate and explain biological function and could spawn better AI.

    “Including such features in computational models may better approximate and explain biological function, which could in turn spawn better AI.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    An additional core ingredient of human intelligence is dynamic information routing that may mirror the packet-switched Internet.

    “it is suggested that an additional”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    The number of Internet hosts increased without major disruption over a few decades, surpassing the number of neurons in the cortex of many primates including the macaque.

    “the number of Internet hostseach of which can potentially communicate with any other within millisecondshas increased without major disruption over a few decades, to surpass the number of neurons in the cortex of many primates including the macaque (Fasolo 2011).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Fasolo 2011

      Not checked in the free run.

  • May combine two claims.

    Premisenone given

    Neural networks commonly use fixed routing, all-to-all connectivity between layers, and feedback only between adjacent layers.

    “in neural networks, one commonly employs a fixed routing system, all-to-all connectivity between layers, and feedback only between adjacent layers.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Emotions are shortcuts and the framework in which cognition happens.

    “Emotions are shortcuts.Emotions are the framework in which cognition happens (e.g., Bach 2009;Dörner 2001).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Bach 2009

      Not checked in the free run.

    • cites Dörner 2001

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    The function of emotions is to adjust behavior to the demands of the current situation.

    “The function of emotions is to adjust behavior to the demands of the current situation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    A packet-switched routing protocol would not necessarily improve current deep learning systems, but may be better suited to modeling approaches that approximate cortical networks.

    “Although a packet-switched routing protocol would not necessarily improve current deep learning systems, it may be better suited to modeling approaches that more closely approximate cortical networks' structure and function.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    These features may make neural networks well suited to learning a particular input space but ill suited to flexible processing and efficient handling of new circumstances.

    “These features may make such systems well suited to learning a particular input space, but ill suited to flexible processing and efficient handling of new circumstances.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisenone given

    Message passing lacks theoretical guarantees but has been known to work well in many practical cases.

    “Despite lacking theoretical guarantees, MP has been known to work well in many practical cases”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given

    Anger means high activation, quick and rough perception, little planning and deliberation, and making a quick choice.

    “Anger means high activation, quick and rough perception, little planning and deliberation, and making a quick choice.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    Emotions can partly explain why humans learn rich models from sparse data.

    “Perhaps emotions can partly explain why humans learn”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning

    A theory or computational model of human intelligence that focuses solely on cognition is not a real theory of human intelligence.

    “In brief, a theory or a computational model of human intelligence that focuses solely on cognition is not a real theory of human intelligence.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    The same meta-learning approach could be used with meta-tasks that are extremely common due to sociocultural conventions.

    “the same approach could be used with meta-tasks that are extremely common as a result of sociocultural conventions, such as”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationreasoning

    Fully addressing the challenges Lake et al. pose is a long-term challenge for the science of learning.

    “Fully addressing the challenges Lake et al. poserather than building in compositional knowledge structures that will ultimately prove limitingis a long-term challenge for the science of learning.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationreasoning

    These errors may be an inevitable part of the learning process by which models are updated based on experience.

    “These errors may be an inevitable part of the learning process by which models are updated based on experience.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisereasoning

    Measuring similarity of processing dynamics between a model and a brain has to rely on summary statistics of activity and may be equally possible for neural and cognitive models.

    “Measuring the similarity of processing dynamics between a model and a brain has to rely on summary statistics of the activity and may be equally possible for neural and cognitive models.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationnone given

    These deficiencies are pervasive and have real-world consequences.

    “These deficiencies are pervasive and have real-world consequences.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing).

    Interpretationcitation

    One could test whether visual representation of characters in high-level visual regions reflects the similarity predicted by Lake et al.'s generative model of character perception.

    “For example, one could test whether the visual representation of characters in high-level visual regions reflects the similarity predicted by the generative model of character perception proposed by Lake et al. (2015a).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Lake et al. 2015a

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Many continue to believe that immunisation increases autism risk, and refusal to immunise has decreased population immunity to preventable diseases.

    “many continue to believe that immunisation increases the risk of autism and their refusal to immunise has decreased the population's immunity to preventable diseases (Larson et al. 2011;Silverman & Hendrix 2015).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Larson et al. 2011

      Not checked in the free run.

    • cites Silverman & Hendrix 2015

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning

    Humanlike cognition should involve the same brain representations and algorithms that the human brain employs.

    “However, humans are biological organisms, and so”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationcitation

    Applying a causal model, especially one based on intuitive understanding, is essentially a gamble.

    “Applying a causal model, especially one based on intuitive understanding, is essentially a gamble.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Humans learn complex skills in a domain after previously learning simpler ones, gradually building structured knowledge.

    “Humans learn complex skills in a domain after previously learning simpler ones, gradually building structured knowledge as they learn.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    As long as major components of human intelligence are out of reach of machines, we are far from understanding the human brain and cognition.

    “As long as major components of human intelligence are out of the reach of machines, we are obviously far from understanding the human brain and cognition.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Avoiding commitment to symbolic compositionality increases flexibility to respond to subtle context influences and allows more robust learning across contexts.

    “Avoiding a commitment to symbolic compositionality increases one's flexibility to respond to sometimes subtle influences of context and allows for the possibility of more robust learning across contexts.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims. The quote only partly states this claim.

    Interpretationreasoning

    One reason humans might learn from fewer games than a neural network trained on chess as a single integrated task is that humans receive feedback on many related tasks throughout learning, allowing more feedback from a single experience and constraining representations.

    “One reason humans might be able to learn from fewer games than a neural network trained on playing chess as a single integrated task is that humans receive feedback on many of these tasks throughout learning, and this both allows more feedback from a single experience”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Premisereasoning

    A model must not only reach human levels of task performance but also predict detailed patterns of behavioral responses.

    “A model must not only reach human levels of task performance, but also predict detailed patterns of behavioral responses (e.g., errors and reaction times on particular instances of a task).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    This reflects that current models learn from scratch.

    “This reflects the fact that current models learn from scratch.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    The meta-tasks used in these first attempts only roughly reflect the training environment humans face.

    “Although the meta-tasks used in these first attempts only roughly reflect the training environment that humans face (we probably do not face explicit one-shot classification problems that frequently)”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Premisereasoning

    A good model should somehow match the brain's dynamics of information processing.

    “A good model should somehow match the brain's dynamics of information processing.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Human knowledge acquisition and generalization can be understood without commitment to domain-specific knowledge structures or compositional knowledge representation.

    “we believe that human knowledge acquisition and generalization can be understood without building in a commitment to domain-specific knowledge structures or compositional knowledge representation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Premisereasoning

    Knowledge of physical relations between objects and psychological relations between agents brings huge advantages.

    “Knowledge of physical relations between objects and psychological relations between agents brings huge advantages.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    It provides transferable information that allows humans to quickly apprehend a new situation.

    “It provides a wealth of transferable information that allows humans to quickly apprehend a new situation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Their inferences require realistically small amounts of data but unrealistically large amounts of computation, and their high-level feats do not always scale to complex real-world challenges.

    “Their inferences require realistically small amounts of data, but unrealistically large amounts of computation, and, as a result, their highlevel feats of cognition do not always scale to complex real-world challenges.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    The brain's inference process appears to involve recurrent cycles of message passing at multiple scales.

    “The brain's inference process appears to involve recurrent cycles of message passing at multiple scales, from local interactions within an area to long-range interactions between higher-and lower-level representations.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given

    Our causal models are often incorrect, resistant to correction, and applied inappropriately to new situations.

    “Our models are often incorrect, resistant to correction, and applied inappropriately to new situations.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning

    This enables humans to infer past causes and future implications, focusing on what matters to behavioral success.

    “This enables us to infer past causes and future implications, with a focus on what matters to our behavioral success.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning

    The interaction of motivation, cognition, and emotion is the real strength of human intelligence compared with artificial intelligence.

    “This interaction of motivation, cognition, and emotion, is the real strength of human intelligence compared with artificial intelligence.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Human learning builds cumulatively on prior learning.

    “This is clearly not the same as human learning, which builds cumulatively on prior learning.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Failures to revise false causal models are common and appear to be an inherent part of human reasoning.

    “Failures to revise false causal models are far from rare.In fact, they seem to be an inherent part of human reasoning.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Lewandowsky et al. 2012

      Not checked in the free run.

    • cites Taylor & Ahn 2012

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Deep convolutional neural networks can rapidly recognize objects as humans can, and their internal representations are remarkably similar to those of the human ventral stream.

    “They can rapidly recognize objects as humans can, and their internal representations are remarkably similar to those of the human ventral stream (Eickenberg et al. 2016;Güçlü & van Gerven 2015;Khaligh-Razavi & Kriegeskorte 2014;Yamins et al. 2014).”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites Eickenberg et al. 2016

      Not checked in the free run.

    • cites Güçlü & van Gerven 2015

      Not checked in the free run.

    • cites Khaligh-Razavi & Kriegeskorte 2014

      Not checked in the free run.

  • May combine two claims.

    Premisenone given

    In games like Chess or Go, human learners can receive feedback on the outcome of an entire game and on individual steps in an action sequence.

    “In games like Chess or Go, human learners can receive feedback not only on the outcome of an entire gamedid the learner succeed or fail?but also on individual steps in an action sequence.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Our causal knowledge is almost always incomplete and sometimes completely wrong.

    “our causal knowledge is almost always incomplete and sometimes completely wrong (Rozenblit & Keil 2002).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Rozenblit & Keil 2002

      Not checked in the free run.

  • May combine two claims.

    Interpretationcitation

    Richer feedback can easily be incorporated into neural networks and can enhance learning.

    “This sort of richer feedback can easily be incorporated into neural networks, and doing so can enhance learning (Gülçehre and Bengio 2016).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Gülçehre and Bengio 2016

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    Even two biological brains of the same species will have different numbers of neurons, and any given neuron may be idiosyncratically specialized and may not have an exact match in the other brain.

    “However, even two biological brains of the same species will have different numbers of neurons, and any given neuron may be idiosyncratically specialized, and may not have an exact match in the other brain.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    In cognitive science, empirical support for models comes from behavioral data.

    “In cognitive science, the empirical support for models comes from behavioral data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning

    Appeals to evolution seem out of place in domains of recent human invention such as character-based writing systems.

    “Appeals to evolution, although they may be plausible for some tasks, seem out of place in domains of recent human invention such as character-based writing systems.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning

    Combining deep neural networks with model-building capacities could bring solutions to some of the world's most pressing problems.

    “combining the computational power of deep-neural networks with model-building capacities could indeed bring solutions to some of the world's most pressing problems.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Humans bring complex knowledge and dynamic models of the world to bear on sensory data.

    “We bring complex knowledge and dynamic models of the world to bear on the sensory data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Premisenone given

    Reverse engineering human intelligence can provide useful guidance for artificial intelligence and machine learning.

    “We agree with the authors' assertion that “reverse engineering human intelligence can usefully inform artificial intelligence and machine learning” (sect.1.1, para.3)”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites sect.1.1, para.3

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Cost-benefit inferences about the Figure 6 scene would probably conclude that one agent wants to kill the other and values the other's death more than life.

    “Cost-benefit-based inferences about situations such as the one depicted in the leftmost picture in Figure 6 of Lake et al. will likely conclude that one agent has a desire to kill the other, and that he or she values higher the state of the other being dead than alive.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    Human decision making typically does not involve rational utility maximization.

    “It has been argued that the process of human decision making does not typically involve rational utility maximization (e.g., Hertwig & Herzog 2009).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Hertwig & Herzog 2009

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationcitation

    Machines that resemble humans can make decisions in ways that are transparent and understandable if they accurately copy human cognitive processes.

    “We argue that human-like machines should be designed to make decisions in transparent and comprehensible ways, which can be achieved by accurately mirroring human cognitive processes.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: AI algorithms draw inspiration from human cognitive systems where humans perform better than machines.

  • May combine two claims.

    Interpretationreasoning

    As-if models based on cost-benefit analysis may be transparent and understandable for simple, non-moral problems, but not always.

    “Although as-if models relying on cost-benefit analysis might be reasonably transparent and comprehensible, for example, when problems are simple and do not involve moral considerations, this might not always be the case.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    In the example, a sheriff and a contract killer share the goal of ending a life, but their different identities would substantially change how humans interpret the scene.

    “In the discussed example, although both a sheriff and a contract killer would have the same goal to end another person's life, the difference in their identity would change the human interpretation in a significant way.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    The proposed intuitive psychology rests on the naïve utility calculus framework.

    “the authors' proposal of an “intuitive psychology” is grounded in the naïve utility calculus framework (Jara-Ettinger et al. 2016)”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Jara-Ettinger et al. 2016

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Ontological confusions are linked to acceptance of alternative medicine, potentially leading to choosing ineffective treatment over evidence-based treatments, sometimes at extreme personal risk.

    “Ontological confusions have been linked to people's acceptance of alternative medicine, potentially leading an individual to choose an ineffective treatment over evidence-based treatments, sometimes at extreme personal risk (Lindeman 2011).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Lindeman 2011

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Naïve utility calculus can infer simple agent-specific variables such as goals and competence.

    “We welcome the applicability of naïve utility calculus for inferring simple information concerning agent-specific variables, such as goals and competence level.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Human causal modelling properties produce pervasive biases and may be symptomatic of the use of mental models rather than limitations on human cognition.

    “Developers of AI systems should proceed with caution, as these properties of human causal modelling produce pervasive biases, and may be symptomatic of the use of mental models rather than the limitations on human cognition.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Heuristics are intended to capture ecologically rational human decision mechanisms.

    “Heuristics not only are meant to capture ecologically rational human decision mechanisms (see Todd & Gigerenzer 2007)”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Todd & Gigerenzer 2007

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    Classifying social situations requires structured cognitive representations.

    “Lake et al. acknowledge this problem and call for structured cognitive representations, which are required for classifying social situations.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Lake et al.

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Effective model-based decision-making is associated with capacities for learning and transferring abstract rules and for cognitive control.

    “For example, effective model-based decision-making is associated with capacities for learning and transferring abstract rules (Don et al. 2016), and for cognitive control (Otto et al. 2015), which may potentially be far more powerful in future AI systems.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Don et al. 2016

      Not checked in the free run.

    • cites Otto et al. 2015

      Not checked in the free run.

295 of 295

  • The paper relies on this without support.

    People are better than machines at solving a range of difficult computational problems including concept learning, scene understanding, language acquisition, language understanding, and speech recognition.

    “Despite recent computational achievements, people are better than machines at solving a range of difficult computational problems, including concept learning, scene understanding, language acquisition, language understanding, speech recognition, and so on.”

  • The paper relies on this without support.

    Children possess the ability and desire to uncover underlying causes of sparsely observed events and use that knowledge to go beyond limited data.

    “Children come with the ability and the desire to uncover the underlying causes of sparsely observed events and to use that knowledge to go far beyond the paucity of the data.”

  • The paper relies on this without support.

    People can comprehend a novel scene in a fraction of a second, or a novel utterance in little more than the time it takes to say and hear it.

    “People can comprehend a novel scene in a fraction of a second, or a novel utterance in little more than the time it takes to say it and hear it.”

  • The paper relies on this without support.

    Once a causal model of a task is learned, humans can use the model to plan action sequences that maximize future reward.

    “Once a causal model of a task has been learned, humans can use the model to plan action sequences that maximize future reward.”

  • The paper relies on this without support.

    Humans combine model-based and model-free learning algorithms both competitively and cooperatively, and these interactions are supervised by metacognitive processes.

    “We review evidence that humans combine model-based and model-free learning algorithms both competitively and cooperatively and that these interactions are supervised by metacognitive processes.”

  • The paper relies on this without support.

    People can integrate across multiple examples of a character to infer optional elements, such as the cross-bar in '7's, and combine variants into a single coherent representation.

    “Furthermore, people can efficiently integrate across multiple examples of a character to infer which have optional elements, such as the horizontal cross-bar in '7's, combining different variants of the same character into a single coherent representation.”

  • The paper relies on this without support.

    The core ingredients of human intelligence are not present in most current learning-based AI systems, and not all of them are present together.

    “they are key building blocks, which are not present in most current learningbased AI systemscertainly not all present together”

  • The paper relies on this without support.

    The DQN starts completely from scratch, without prior experience.

    “The DQN is starting completely from scratch”

  • The authors state this as an assumption.

    The success of the BPL framework in the Characters Challenge can be generalized to understanding human learning as rapid model building.

    “The following sections discuss the three main ingredientscompositionality, causality, and learning-to-learnthat were important to the success of this framework and, we believe, are important to understanding human learning as rapid model building more broadly.”

  • The paper relies on this without support.

    The Characters Challenge and the Frostbite Challenge are representative testbeds for studying human-like learning and thinking.

    “We can see this potential in both of our challenge problems.”

  • The authors state this as an assumption.

    Cognitive plausibility offers a surer foundation than biological plausibility for theories of intelligence.

    “For now, we believe cognitive plausibility offers a surer foundation.”

  • The paper relies on this without support.

    Brain cells are not motile enough to selectively deliver messages to their destinations.

    “Given that brain cells themselves are not motile enough to selectively deliver messages to their destination (unlike cells in the immune system, for example)”

  • The paper relies on this without support.

    The brain implements a generative model.

    “what kind of generative model the brain implements”

  • The paper relies on this without support.

    The power of neural networks lies in their ability to encode rich featural and contextual information.

    “On the other hand, the power of neural networks lies precisely in their ability to encode rich featural and contextual information.”

  • The paper relies on this without support.

    Experience with particular causal relations likely correlates with experience of more superficial features.

    “Given that experience with particular causal relations is likely to correlate with experience of more superficial features, a more powerful AI model generator may still suffer similar problems when faced with the difficult decision of which model to apply to a new situation.”

  • The paper relies on this without support.

    Current deep learning research does not include physical embodiment.

    “and yet it is out of the picture in current deep learning research”

  • The paper relies on this without support.

    Deep learning architectures often concentrate on inference and optimization.

    “Deep learning architectures often focus on inference and optimization.”

  • The paper relies on this without support.

    Learning happens through complex dynamical interactions among inference, memory, attention, motivation, low-level sensorimotor loops, embodiment, and social interaction.

    “learning occurs through complex dynamical interaction among systems of inference, memory, attention, motivation, low-level sensorimotor loops, embodiment, and social interaction”

  • The paper relies on this without support.

    Attention and memory are among the ingredients already present in current deep learning research.

    “Although some of these ingredients are part of current DL research, (e.g., attention and memory)”

  • The paper relies on this without support.

    AI algorithms draw inspiration from human cognitive systems where humans perform better than machines.

    “Artificial intelligence algorithms seek inspiration from human cognitive systems in areas where humans outperform machines.”

  • The paper relies on this without support.

    A human-like classification of the Figure 6 scene would conclude that it shows either a legal execution or a murder.

    “a human-like classification of such a scene would rather reach the conclusion that the scene depicts either a legal execution or a murder”

  • The paper relies on this without support.

    People interacting with the system will probably want a justification for why it chose one interpretation over another.

    “Humans interacting with the system will likely expect a justification of why it has picked one interpretation rather than another”
22 of 22

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Models trained with self-supervision can capture the underlying structures and relationships behind their training data.

    “Specifically, we know that models trained with self-supervision can capture the underlying structures and relationships behind their training data [51] -exactly what we need from a compression heuristic.”

    From Learning to (Learn at Test Time): RNNs with Expressive Hidden States (2024) · cites this paper as [51]

    Section 4.2, “Learning as rapid model building” (page unavailable), says unsupervised learning adjusts a model’s internal patterns to match the statistics of input data. That supports a narrower claim about learning from data, but our copy does not establish that self-supervised models capture the underlying structures and relationships behind their training data. (LLM, conf=0.73)

    Experimental — accuracy not yet measured

  • Bayesian inference and evidence accumulation are basic unconscious mechanisms for humans.

    “Bayesian inference and evidence accumulation, which are cornerstone computations for AI ( 2 ), are basic unconscious mechanisms for humans.”

    From What is consciousness, and could machines have it? (2017) · cites this paper as [2]

    The citation is attached to the claim that Bayesian inference and evidence accumulation are cornerstone AI computations, not to the sentence’s claim about unconscious human mechanisms. Section 4.3.1, “Approximate inference in structured models” (page unavailable), discusses Bayesian models and inference in machine learning, but the passages found do not establish evidence accumulation as a cornerstone AI computation. (LLM, conf=0.82)

    Experimental — accuracy not yet measured

  • Bayesian approaches to AI have recognized the usefulness of learning generative models that can be jointly used for actual perception, prospective planning, and retrospective analysis.

    “Bayesian approaches to AI ( 2 , 73 ) have recognized the usefulness of learning generative models that can be jointly used for actual perception (present), prospective planning (future), and retrospective analysis (past).”

    From What is consciousness, and could machines have it? (2017) · cites this paper as [2]

    Section 4.2, “Learning as rapid model building” (page unavailable), links causal generative models to perception and describes models of environmental transitions for control. Section 4.1, “Developmental start-up software” (page unavailable), says simulated planning processes can predict what an agent might do next and support inverse inference from observed actions. These passages support related uses of models, but do not establish that a learned generative model is jointly used for perception, prospective planning, and retrospective analysis. (LLM, conf=0.79)

    Experimental — accuracy not yet measured

  • LLMs do not possess true empathy or social understanding.

    “without possessing any true empathy or social understanding”

    From The benefits and dangers of anthropomorphic conversational agents (2025) · cites this paper as [63]

    In “Thinking like animals or thinking like colleagues?” (page not supplied), the paper says existing AIs lack genuine interests even if they can communicate in natural language. The Kriegeskorte and Mok commentary (page not supplied) says most artificial agents lack the basis for psychological competence. These passages support the general concern about empathy and social understanding, but do not establish the categorical claim about LLMs specifically. (LLM, conf=0.82)

    Experimental — accuracy not yet measured

  • LLMs' inner workings are fundamentally different to human cognition, and they are incapable of feeling or understanding emotion or language in any genuine sense.

    “it remains obviously true that LLMs' inner workings are fundamentally different to human cognition (63), and that they are incapable of feeling or even understanding emotion or language in any genuine sense (27, 64).”

    From The benefits and dangers of anthropomorphic conversational agents (2025) · cites this paper as [63]

    The discussion in §5.1 says deep-learning models have yet to incorporate core ingredients of human-like learning and thought, and §5.3 says neural networks are far from implementing human language abilities. These passages support a difference between the systems studied and human cognition, but do not establish the claim specifically for LLMs. This citation is attached to the first clause; the separate claim about feeling and understanding is cited to other sources. (LLM, conf=0.84)

    Experimental — accuracy not yet measured

Cites

Flagged first, then by citations

Research works printed but not linked: 40

  1. 1961, GPS, A program that simulates human thought
  2. 1966, Cognitive psychology
  3. 2012, JMLR Workshop on Unsupervised and Transfer Learning, 27, 195
  4. 2007, Contemporary Educational Psychology, 25, 54
  5. 1996, Rethinking innateness
  6. 2003, Artificial intelligence: A modern approach
  7. 2015, Journal of Experimental & Theoretical Artificial Intelligence, 28, 751
  8. 2014, Proceedings of Machine Learning Research, 32, 647
  9. 1986, Machine Learning, 1, 47
  10. 1986, Parallel distributed processing: Explorations in the microstructure of cognition, Vol. 1
  11. 1986, Parallel distributed processing: Explorations in the microstructure of cognition, Vol. 1, 216
  12. 2004, Semantic cognition
  13. 2015, Communications of the ACM, 58, 92
  14. 1986, Parallel distributed processing: Explorations in the microstructure of cognition, Vol. 2
  15. 1995, An invitation to cognitive science: vol. 2. Visual cognition, 297
  16. Ghahramani, 2014, Advances in neural information processing systems 27 (NIPS 2014)
  17. 1979, Cognitive Science, 3, 275
  18. 1950, Mind, 59, 433
  19. 2007, Nature, 450, 57
  20. 2010, Origins of human communication
  21. 1962, IRE Transactions on Information Theory, 8, 155
  22. 1982, Vision
  23. 2003, Journal of Experimental Psychology: Learning, Memory, and Cognition, 29, 1141
  24. 2007, The stuff of thought: Language as a window into human nature
  25. 1989, Categorization and naming in children
  26. 1978, Papers and Reports on Child Language Development, 15, 17
  27. 1998, Cognitive Psychology, 282, 243
  28. Bresnan, 1978, Linguistic theory and psychological reality, 264
  29. Woodward, 2009, Learning and the infant mind, 66
  30. 2006, Mind as machine: A history of cognitive science
  31. 2014, Machine: earning: An artificial intelligence approach, vol. 3
  32. 2014, PLoS Computational Biology, 10
  33. 1990, Technometric, 33, 251
  34. 1975, The psychology of computer vision, 157
  35. Damon, 1998, Handbook of child psychology: Vol. 2. Cognition, perception, and language development, 523
  36. 1992, Machine Learning, 8, 279
  37. 2012, IEEE Signal Processing Magazine, 29, 82
  38. 2013, Frontiers in Psychology, 4, 365
  39. 2015, Science, 34, 273
  40. 2004, Bayesian data analysis

Cited by

Sorted by citations

Related

Authors
Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J.
Affiliations
  • New York University (US)
  • Massachusetts Institute of Technology (US)
  • Harvard University (US)
ORCID
Venue
Behavioral and Brain Sciences
Publisher
Cambridge University Press (CUP)
Published
2016-11-24
Volume
40
Pages
e253-e253
ISSN
  • 0140-525X (ISSN-L)
  • 1469-1825
DOI
10.1017/s0140525x16001837
OpenAlex
W2328078142
Cited by
2,550
References
274
Citations by year
2026: 244 · 2025: 273 · 2024: 305 · 2023: 274 · 2022: 253 · 2021: 358 · 2020: 282 · 2019: 280 · 2018: 168 · 2017: 90 · 2016: 19
FWCI
152.60
Type
journal article
Language
en
Open access
Licence
www.cambridge.org/core/terms (unspecified)
Locations
  • Behavioral and Brain Sciences (publishedVersion)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • DSpace@MIT (Massachusetts Institute of Technology) (submittedVersion, cc-by-nc-sa)
  • arXiv (Cornell University)
Funders
  • National Science Foundation (ES/K002201/1, RP2012-V-022, 1231216)
  • York University
  • Leverhulme Trust (RP2012-V022, ES/K002201/1)
  • Research Councils UK (EP/K039830/1)
  • Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung (146702, 144413)
  • Companhia Brasileira de Metalurgia e Mineração (CCF-1231216)
  • Economic and Social Research Council (ES/K002201/1, RP2012-V-022, K002201, ES/K002201/)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Psychiatry, Mental Health, Neuroscience
  • Physical Sciences › Computer Science › Artificial Intelligence › AI-based Problem Solving and Planning
  • Physical Sciences › Computer Science › Artificial Intelligence › Cognitive Computing and Networks
Keywords
human-like learning, artificial intelligence, causal models, explanation and understanding, intuitive theories, compositionality, meta-learning, knowledge generalization, cognitive models, deep neural network
MeSH terms
Achievement, Artificial Intelligence, Humans, Intelligence, Thinking, Visual Perception, Neural Networks, Computer

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-05.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-05.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review