Toward an integration of deep learning and neuroscience

Not all learning requires a general-purpose optimization mechanism like gradient descent.

“Not all learning requires a general-purpose optimization mechanism like gradient descent5.”

A range of implementations of credit assignment through multiple layers of neurons are compatible with current knowledge of neural circuitry.

“we argue that a range of implementations of credit assignment through multiple layers of neurons are compatible with our current knowledge of neural circuitry”

In an appropriate network context, STDP learning could give rise to a biological implementation of backpropagation.

“In other words, in an appropriate network context, STDP learning could give rise to a biological implementation of backpropagation10.”

• cites Xie and Seung, 2000Not checked yet

  Not checked in the free run.

• cites Bengio et al., 2015bNot checked yet

  Not checked in the free run.

Local self-organization mechanisms are likely insufficient to account for the brain's powerful learning performance.

“We argue that the above mechanisms of local self-organization are likely insufficient to account for the brain's powerful learning performance (Brea and Gerstner, 2016).”

• cites Brea and Gerstner, 2016Not checked yet

  Not checked in the free run.

The brain's specialized systems can be interpreted as enabling efficient optimization for specific problem classes.

“the brain's specialized systems can be interpreted as enabling efficient optimization for specific problem classes”

The biological implementations of backpropagation proposed above do not give a natural implementation of backpropagation through time for recurrent networks.

“The biological implementations of backpropagation proposed above, while applicable to feedforward networks, do not give a natural implementation of “backpropagation through time” (BPTT) (Werbos, 1990) for recurrent networks”

• cites Werbos, 1990 sourceNot checked yet

  Not checked in the free run.

The brain uses several different types of optimization to solve distinct problems.

“We argue that the brain uses several different types of optimization to solve distinct problems.”

The brain uses cost functions as a key driving force of its development.

“Thus, the brain uses cost functions as a key driving force of its development, much as modern machine learning systems do.”

Feedback alignment is more a primitive that can be used in fully biological approximate implementations of backpropagation than a fully biological implementation in its own right.

“Thus, feedback alignment is more a primitive that can be used in fully biological (approximate) implementations of backpropagation, than a fully biological implementation in its own right.”

• cites Lillicrap et al., 2014Not checked yet

  Not checked in the free run.

• cites Liao et al., 2015Not checked yet

  Not checked in the free run.

There are now many proposed biologically plausible mechanisms by which a neural circuit could implement optimization algorithms that, like backpropagation, can efficiently make use of the gradient.

“Contrary to what was once a common assumption, there are now many proposed “biologically plausible” mechanisms by which a neural circuit could implement optimization algorithms that, like backpropagation, can efficiently make use of the gradient.”

• cites O'Reilly, 1996 sourceNot checked yet

  Not checked in the free run.

• cites Xie and Seung, 2003Not checked yet

  Not checked in the free run.

• cites Lillicrap et al., 2014Not checked yet

  Not checked in the free run.

Individual neurons should not be regarded as single nodes but as multi-component sub-networks.

“Individual neurons thus should not be regarded as single “nodes” but as multi-component sub-networks.”

The types of bi-directional, non-linear, multi-variate interactions that are possible inside a single neuron could support gradient descent learning or other powerful optimization mechanisms.

“But it is clear that the types of bi-directional, non-linear, multi-variate interactions that are possible inside a single neuron could support gradient descent learning or other powerful optimization mechanisms.”

One way of allowing networks to have one-shot properties is I-theory, in the context of learning invariant representations for object recognition.

“One way of allowing networks to have such properties is what is described by I-theory, in the context of learning invariant representations for object recognition (Anselmi et al., 2015).”

• cites Anselmi et al., 2015 sourceNot checked yet

  Not checked in the free run.

The backpropagation of action potentials into active dendritic branches potentially simplifies the problem of credit assignment.

“However, it propagates more strongly into the branches of the dendritic tree that have been active (Williams and Stuart, 2000), potentially simplifying the problem of credit assignment (Körding and König, 2000).”

• cites Körding and König, 2000Not checked yet

  Not checked in the free run.

Neuromodulation seems to expand the range of possible algorithms that could be used for optimization.

“Overall, neuromodulation seems to expand the range of possible algorithms that could be used for optimization.”

If the system is given access to appropriate memory stores and representations of temporal context, this could mitigate the need for temporal credit assignment as such.

“If the system is given access to appropriate memory stores and representations (Buonomano and Merzenich, 1995; Gershman et al., 2012, 2014) of temporal context, this could potentially mitigate the need for temporal credit assignment as such”

• cites Buonomano and Merzenich, 1995Not checked yet

  Not checked in the free run.

• cites Gershman et al., 2012 sourceNot checked yet

  Not checked in the free run.

• cites Gershman et al., 2014 sourceNot checked yet

  Not checked in the free run.

Spiking recurrent networks using realistic population coding schemes can, with an appropriate choice of connection weights, compute complicated, cognitively relevant functions.

“In any case, there is little doubt that spiking recurrent networks using realistic population coding schemes can, with an appropriate choice of connection weights, compute complicated, cognitively relevant functions17.”

Neuromodulation is another potentially important biological primitive.

“Another potentially important biological primitive is neuromodulation.”

Modulators can be used to gate synaptic plasticity on and off selectively in different areas and at different times, allowing precise, rapidly updated orchestration of where and when cost functions are applied.

“First, modulators can be used to gate synaptic plasticity on and off selectively in different areas and at different times, allowing precise, rapidly updated orchestration of where and when cost functions are applied.”

There is the possibility for a neuron to store both its activation itself, and the error derivative of a cost function with respect to its activation, as required in backpropagation.

“Thus, there is the possibility for a neuron to store both its activation itself, and the error derivative of a cost function with respect to its activation, as required in backpropagation, and biological implementations of backpropagation based on this principle have been proposed (Körding and König, 2001; Schiess et al., 2016)20.”

The same neuron or circuit can exhibit different input-output responses and plasticity depending on a global circuit state, as reflected by the concentrations of various neuromodulators.

“The same neuron or circuit can exhibit different input-output responses and plasticity depending on a global circuit state, as reflected by the concentrations of various neuromodulators like dopamine, serotonin, norepinephrine, acetylcholine, and hundreds of different neuropeptides such as opiods (Bargmann, 2012; Bargmann and Marder, 2013).”

• cites Bargmann, 2012 sourceNot checked yet

  Not checked in the free run.

• cites Bargmann and Marder, 2013 sourceNot checked yet

  Not checked in the free run.

This replay would allow the brain to move an item from episodic memory into a long-term memory stored in the weights of a cortical network, by replaying the memory over and over.

“This would allow the brain to move an item from episodic memory into a long-term memory stored in the weights of a cortical network (Ji and Wilson, 2007), by replaying the memory over and over.”

• cites Ji and Wilson, 2007 sourceNot checked yet

  Not checked in the free run.

There are population coding mechanisms that could support explicit probabilistic computations.

“Moreover, there are population coding mechanisms that could support explicit probabilistic computations (Zemel and Dayan, 1997; Sahani and Dayan, 2003; Rao, 2004; Ma et al., 2006; Eliasmith and Martens, 2011; Gershman and Beck, 2016).”

• cites Zemel and Dayan, 1997Not checked yet

  Not checked in the free run.

• cites Sahani and Dayan, 2003Not checked yet

  Not checked in the free run.

• cites Rao, 2004 sourceNot checked yet

  Not checked in the free run.

These potential substrates of one-shot learning rely on mechanisms other than simple gradient descent.

“These potential substrates of one-shot learning rely on mechanisms other than simple gradient descent.”

Diverse mechanisms like retrograde signals using cannabinoids, or rapidly-diffusing gases such as nitric oxide, are among many that could enable learning rules that go beyond conventional conceptions of backpropagation.

“Beyond dendritic computation, diverse mechanisms (Marblestone and Boyden, 2014) like retrograde (post-synaptic to pre-synaptic) signals using cannabinoids (Wilson and Nicoll, 2001), or rapidly-diffusing gases such as nitric oxide (Arancio et al., 1996), are among many that could enable learning rules that go beyond conventional conceptions of backpropagation.”

• cites Marblestone and Boyden, 2014Not checked yet

  Not checked in the free run.

• cites Wilson and Nicoll, 2001Not checked yet

  Not checked in the free run.

• cites Arancio et al., 1996 sourceNot checked yet

  Not checked in the free run.

Dendritic computation is a biological primitive that could impact prospects for learning algorithms in several ways.

“One such biological primitive is dendritic computation, which could impact prospects for learning algorithms in several ways.”

A primary function of the cortex is some form of unsupervised learning via prediction.

“These models generally agree that a primary function of the cortex is some form of unsupervised learning via prediction (O'Reilly et al., 2014b; Brea et al., 2016)22.”

• cites O'Reilly et al., 2014bNot checked yet

  Not checked in the free run.

• cites Brea et al., 2016Not checked yet

  Not checked in the free run.

This solution effectively uses many iterations of weight updating to fully learn a single item, even if one has only been exposed to it once.

“This solution effectively uses many iterations of weight updating to fully learn a single item, even if one has only been exposed to it once.”

Biological implementations of backpropagation based on the principle of neurons storing multiple variables have been proposed.

“Thus, there is the possibility for a neuron to store both its activation itself, and the error derivative of a cost function with respect to its activation, as required in backpropagation, and biological implementations of backpropagation based on this principle have been proposed (Körding and König, 2001; Schiess et al., 2016)20.”

• cites Körding and König, 2001Not checked yet

  Not checked in the free run.

• cites Schiess et al., 2016 sourceNot checked yet

  Not checked in the free run.

Stochastic gradient descent can be used to generate a system that samples adaptively.

“Interestingly, however, stochastic gradient descent can be used to generate a system that samples adaptively (Alain et al., 2015; Bouchard et al., 2015).”

• cites Alain et al., 2015Not checked yet

  Not checked in the free run.

• cites Bouchard et al., 2015Not checked yet

  Not checked in the free run.

These modulators interact in complex and cell-type-specific ways to influence circuit function.

“These modulators interact in complex and cell-type-specific ways to influence circuit function.”

• cites Bargmann, 2012 sourceNot checked yet

  Not checked in the free run.

• cites Bargmann and Marder, 2013 sourceNot checked yet

  Not checked in the free run.

Standard gradient descent does not incorporate any adaptive sampling mechanism, e.g., it does not deliberately sample data so as to maximally reduce its uncertainty.

“Standard gradient descent does not incorporate any such adaptive sampling mechanism, e.g., it does not deliberately sample data so as to maximally reduce its uncertainty.”

The nonlinear and dynamical properties of cortical pyramidal neurons, especially their multiple dendritic zones targeted by different kinds of projections, may allow the pyramidal neuron to make comparisons of top-down and bottom-up inputs.

“For example, the nonlinear and dynamical properties of cortical pyramidal neurons—the principal excitatory neuron type in cortex (Shepherd, 2014)—are of particular interest here, especially because these neurons have multiple dendritic zones that are targeted by different kinds of projections, which may allow the pyramidal neuron to make comparisons of top-down and bottom-up inputs23.”

• cites Shepherd, 2014 sourceNot checked yet

  Not checked in the free run.

In I-theory, instead of training via gradient descent, image templates are stored in the weights of simple-complex cell networks while objects undergo transformations, similar to the use of stored templates in HMAX.

“Instead of training via gradient descent, image templates are stored in the weights of simple-complex cell networks while objects undergo transformations, similar to the use of stored templates in HMAX (Serre et al., 2007).”

• cites Serre et al., 2007 sourceNot checked yet

  Not checked in the free run.

The existence of these diverse mechanisms may call into question traditional, intuitive notions of biological plausibility for learning algorithms.

“The existence of these diverse mechanisms may call into question traditional, intuitive notions of “biological plausibility” for learning algorithms.”

The brain may have discovered mechanisms of credit assignment quite different from those in machine learning.

“This suggests that the brain may have discovered mechanisms of credit assignment quite different from those dreamt up by machine learning.”

The complex physiology of individual biological neurons may provide mechanisms for learning that go beyond backpropagation.

“The complex physiology of individual biological neurons may not only help explain how some form of efficient gradient descent could be implemented within the brain, but also could provide mechanisms for learning that go beyond backpropagation.”

The nervous system may have a way of quickly storing and replaying sequences of events.

“Additionally, the nervous system may have a way of quickly storing and replaying sequences of events.”

Higher-level systems in the brain may be able to implement Bayesian learning of sequential programs, which is a powerful means of one-shot learning.

“Finally, higher-level systems in the brain may be able to implement Bayesian learning of sequential programs, which is a powerful means of one-shot learning (Lake et al., 2015).”

• cites Lake et al., 2015 sourceNot checked yet

  Not checked in the free run.

Recent architectural advances, including specialized spatial attention and feedback mechanisms, as well as specialized memory mechanisms, do allow some types of one-shot generalization to be driven by backpropagation-based learning.

“It should be noted, though, that recent architectural advances, including specialized spatial attention and feedback mechanisms (Rezende et al., 2016), as well as specialized memory mechanisms (Santoro et al., 2016), do allow some types of one-shot generalization to be driven by backpropagation-based learning.”

• cites Rezende et al., 2016Not checked yet

  Not checked in the free run.

• cites Santoro et al., 2016Not checked yet

  Not checked in the free run.

Specific spike-timing-dependent and homeostatic plasticity rules have been shown to give rise to gradient descent on a prediction error in recurrent neural networks.

“For example, specific spike-timing-dependent and homeostatic plasticity rules have been shown to give rise to gradient descent on a prediction error in recurrent neural networks (Galtier and Wainrib, 2013).”

• cites Galtier and Wainrib, 2013Not checked yet

  Not checked in the free run.

For reinforcement learning of a deep network to be biologically plausible, we need a more powerful learning mechanism, since we are learning based on a more limited evaluation signal than in the supervised case: we do not have the full target pattern to train toward.

“For reinforcement learning of a deep network to be biologically plausible, however, we need a more powerful learning mechanism, since we are learning based on a more limited evaluation signal than in the supervised case: we do not have the full target pattern to train toward.”

A system can learn, by gradient descent, how to choose its own input data samples in order to learn most quickly from them by gradient descent.

“In other words, a system can learn, by gradient descent, how to choose its own input data samples in order to learn most quickly from them by gradient descent.”

This process for generating learning targets provides a kind of bridge between reinforcement learning and efficient backpropagation-based gradient descent learning.

“This process for generating learning targets provides a kind of bridge between reinforcement learning and efficient backpropagation-based gradient descent learning29.”

In such systems, the objective function and the optimization mechanism can thus be mixed and matched modularly.

“In such systems, the objective function and the optimization mechanism can thus be mixed and matched modularly.”

A key remaining issue in reinforcement learning is the problem of long timescales, e.g., learning the many small steps needed to navigate from London to Chicago.

“Even given these advances, a key remaining issue in reinforcement learning is the problem of long timescales, e.g., learning the many small steps needed to navigate from London to Chicago.”

Consequently, reinforcement learning is generally much less efficient than supervised learning.

“Consequently, reinforcement learning is generally much less efficient than supervised learning.”

Many of the formal guarantees of reinforcement learning suggest that the difference between an optimal policy and the learned policy becomes increasingly loose as the discount factor shifts to take into account reward at longer timescales.

“Many of the formal guarantees of reinforcement learning (Williams and Baird, 1993), for example, suggest that the difference between an optimal policy and the learned policy becomes increasingly loose as the discount factor shifts to take into account reward at longer timescales.”

• cites Williams and Baird, 1993Not checked yet

  Not checked in the free run.

Unsupervised pre-training of deep networks is an instance of this, typically making use of a stacked auto-encoder framework.

“Unsupervised pre-training of deep networks is an instance of this (Erhan and Manzagol, 2009), typically making use of a stacked auto-encoder framework.”

• cites Erhan and Manzagol, 2009Not checked yet

  Not checked in the free run.

In many cases, the system may be more limited by the available training signals than by the optimization process itself.

“But in many cases, the system may be more limited by the available training signals than by the optimization process itself.”

The feedback pathway, together with a diffusible reward signal, together gate plasticity.

“The feedback pathway, together with a diffusible reward signal, together gate plasticity.”

Specific unsupervised objectives could be implemented implicitly through specific local network dynamics and plasticity rules inside a network without explicit computation of cost function, nor explicit propagation of error derivatives.

“Thus, specific unsupervised objectives could be implemented implicitly through specific local network dynamics32 and plasticity rules inside a network without explicit computation of cost function, nor explicit propagation of error derivatives.”

Machine learning has not yet developed methods to deal with such a wide range of timescales and scopes of hierarchical action.

“Machine learning has not yet developed methods to deal with such a wide range of timescales and scopes of hierarchical action.”

It is possible to reinforcement-learn a policy for selecting the most interesting inputs to drive learning.

“In other words, it is possible to reinforcement-learn a policy for selecting the most interesting inputs to drive learning.”

The AGREL networks and variants like KickBack give a way to compute an approximation to the full gradient in a reinforcement learning context using a feedback-based attention mechanism for credit assignment within the multi-layer network.

“The “attention-gated reinforcement learning” (AGREL) networks of Stanisor et al. (2013), Brosch et al. (2015), and Roelfsema and van Ooyen (2005), and variants like KickBack (Balduzzi, 2014), give a way to compute an approximation to the full gradient in a reinforcement learning context using a feedback-based attention mechanism for credit assignment within the multi-layer network.”

• cites Stanisor et al., 2013 sourceNot checked yet

  Not checked in the free run.

• cites Brosch et al., 2015 sourceNot checked yet

  Not checked in the free run.

• cites Roelfsema and van Ooyen, 2005Not checked yet

  Not checked in the free run.

Hence, some form of trial and error twiddling must be used to discover how to increase the evaluation signal.

“Hence, some form of trial and error twiddling must be used to discover how to increase the evaluation signal.”

These approaches achieve optimality in an arbitrary, abstract environment.

“These approaches achieve optimality in an arbitrary, abstract environment.”

The cost function used to train a given sub-network in the brain is a key innate property that can be built into the system by evolution.

“Thus, the cost function used to train a given sub-network in the brain is a key innate property that can be built into the system by evolution.”

In one common form of unsupervised learning, higher brain areas attempt to produce samples that are statistically similar to those actually seen in lower layers.

“In one common form of unsupervised learning, higher brain areas attempt to produce samples that are statistically similar to those actually seen in lower layers.”

• cites Hinton et al., 1995 sourceNot checked yet

  Not checked in the free run.

In unsupervised learning, cost functions may not take the form of externally supplied training or error signals, but rather can be built into the dynamics inherent to the network itself, i.e., there may be no need for a separate circuit to compute and impose a cost function on the network.

“In unsupervised learning, cost functions may not take the form of externally supplied training or error signals, but rather can be built into the dynamics inherent to the network itself, i.e., there may be no need for a separate circuit to compute and impose a cost function on the network.”

Deep Q-learning demonstrates reinforcement learning in a deep network, wherein most of the network is trained via backpropagation.

“Deep Q-learning (Mnih et al., 2015) demonstrates reinforcement learning in a deep network, wherein most of the network is trained via backpropagation.”

• cites Mnih et al., 2015 sourceNot checked yet

  Not checked in the free run.

Recently, combining such prediction-based learning with a specific gating mechanism has been shown to lead to unsupervised learning of disentangled representations.

“Recently, combining such prediction-based learning with a specific gating mechanism has been shown to lead to unsupervised learning of disentangled representations (Whitney et al., 2016).”

• cites Whitney et al., 2016Not checked yet

  Not checked in the free run.

A range of unsupervised machine learning techniques, such as the sparse autoencoders used to discover cats in YouTube videos, build sparseness into neural networks.

“A range of unsupervised machine learning techniques, such as the sparse autoencoders (Le et al., 2012) used to discover cats in YouTube videos, build sparseness into neural networks.”

• cites Le et al., 2012Not checked yet

  Not checked in the free run.

Each sensory modality is primarily connected to one part of the brain.

“Each sensory modality is primarily connected to one part of the brain36.”

Object locations and configurations tend to be coherent in time.

“Moreover, object locations and configurations tend to be coherent in time.”

The network can thus implement a statistical goal that embodies some understanding of the way that the world produces samples.

“The network can thus implement a statistical goal that embodies some understanding of the way that the world produces samples33.”

This idea of continuity is used in a great number of artificial neural networks and related models.

“This idea of continuity is used in a great number of artificial neural networks and related models (Wiskott and Sejnowski, 2002; Földiák, 2008; Mobahi et al., 2009).”

• cites Wiskott and Sejnowski, 2002Not checked yet

  Not checked in the free run.

• cites Földiák, 2008 sourceNot checked yet

  Not checked in the free run.

• cites Mobahi et al., 2009 sourceNot checked yet

  Not checked in the free run.

As young babies, we already know that objects tend to persist over time.

“As young babies, we already know (Bremner et al., 2015) that objects tend to persist over time.”

• cites Bremner et al., 2015 sourceNot checked yet

  Not checked in the free run.

Unsupervised learning mechanisms that maximize temporal coherence or slowness are increasingly used in machine learning.

“Unsupervised learning mechanisms that maximize temporal coherence or slowness are increasingly used in machine learning35.”

Higher levels of cortex in each modality are heavily connected to the other modalities.

“But higher levels of cortex in each modality are heavily connected to the other modalities.”

Then, no special learning rules beyond simple backpropagation are needed.

“Then, no special learning rules beyond simple backpropagation are needed.”

Neural networks can also be designed to learn to invert spatial transformations.

“Neural networks can also be designed to learn to invert spatial transformations (Jaderberg et al., 2015).”

• cites Jaderberg et al., 2015Not checked yet

  Not checked in the free run.

The output of an object recognition system must have sparse activations.

“As such, we know that the output of an object recognition system must have sparse activations.”

Biologically plausible theories are available for how cortex could learn using prediction errors by exploiting temporal differences or top-down feedback.

“and biologically plausible theories are available for how cortex could learn using prediction errors by exploiting temporal differences (O'Reilly et al., 2014b) or top-down feedback (George and Hawkins, 2009).”

• cites O'Reilly et al., 2014bNot checked yet

  Not checked in the free run.

• cites George and Hawkins, 2009 sourceNot checked yet

  Not checked in the free run.

Through either explicitly statistical or gradient descent based learning, the brain can thus obtain a probabilistic model that simulates features of the world.

“Through either explicitly statistical or gradient descent based learning, the brain can thus obtain a probabilistic model that simulates features of the world.”

The same is true about multiple parts of the same modality: if we understand the left half of the visual field, it tells us an awful lot about the right.

“The same is true about multiple parts of the same modality: if we understand the left half of the visual field, it tells us an awful lot about the right.”

Each modality effectively produces training signals for the others.

“This way, each modality effectively produces training signals for the others37.”

Evidence from psychophysics suggests that some kind of training via detection of sensory conflicts may be occurring in children.

“Evidence from psychophysics suggests that some kind of training via detection of sensory conflicts may be occurring in children (Nardini et al., 2010).”

• cites Nardini et al., 2010 sourceNot checked yet

  Not checked in the free run.

Ladder autoencoders make use of lateral connections and local noise injection to introduce an unsupervised cost function, based on internal reconstructions, that can be readily combined with supervised cost functions defined on the networks top layer outputs.

“Ladder autoencoders make use of lateral connections and local noise injection to introduce an unsupervised cost function, based on internal reconstructions, that can be readily combined with supervised cost functions defined on the networks top layer outputs (Valpola, 2015).”

• cites Valpola, 2015 sourceNot checked yet

  Not checked in the free run.

Learning rules for generative models have historically involved local message passing of a form quite different from backpropagation, e.g., in a multi-stage process that first learns one layer at a time and then fine-tunes via the wake-sleep algorithm.

“Learning rules for generative models have historically involved local message passing of a form quite different from backpropagation, e.g., in a multi-stage process that first learns one layer at a time and then fine-tunes via the wake-sleep algorithm (Hinton et al., 2006).”

• cites Hinton et al., 2006 sourceNot checked yet

  Not checked in the free run.

In one implementation, a system can simply predict the next input delivered to the system and can then use the difference between the actual next input and the predicted next input as a full vectorial error signal for supervised gradient descent.

“In one implementation, a system can simply predict the next input delivered to the system and can then use the difference between the actual next input and the predicted next input as a full vectorial error signal for supervised gradient descent.”

Building in such spatio-temporal sparseness priors should serve as an inductive bias that can accelerate learning.

“Building in such spatio-temporal sparseness priors should serve as an “inductive bias” (Mitchell, 1980) that can accelerate learning.”

• cites Mitchell, 1980Not checked yet

  Not checked in the free run.

There are multiple modalities of input to the brain.

“Furthermore, there are multiple modalities of input to the brain.”

We can formulate this prior knowledge as a cost function, for example by penalizing representations which are not temporally continuous.

“We can formulate this prior knowledge as a cost function, for example by penalizing representations which are not temporally continuous.”

Generative models can also be trained via backpropagation.

“Generative models can also be trained via backpropagation.”

The emergence or disappearance of an object from a region of space is a rare event.

“The emergence or disappearance of an object from a region of space is a rare event.”

Having the right kind of algorithm and data structure in place to solve a problem allows it to be solved efficiently, robustly and with a minimum amount of learning or optimization needed.

“Having the right kind of algorithm and data structure in place to solve a problem allows it to be solved efficiently, robustly and with a minimum amount of learning or optimization needed.”

These mechanisms may need to be at least partially in place in advance of learning.

“These mechanisms may need to be at least partially in place in advance of learning.”

To enable the learning of efficient computational solutions, the brain may need pre-specialized systems for planning and executing sequential multi-step processes, for accessing memories, and for forming and manipulating compositional and recursive structures.

“In particular, to enable the learning of efficient computational solutions, the brain may need pre-specialized systems for planning and executing sequential multi-step processes, for accessing memories, and for forming and manipulating compositional and recursive structures40.”

The cortex may have evolved as a trainable module for which the training is orchestrated by these older structures.

“For example, the cortex may have evolved as a trainable module for which the training is orchestrated by these older structures.”

Cost functions can be specified that could lead to the learning of increasingly sophisticated mental abilities in a biologically plausible manner.

“In this section, we have seen how cost functions can be specified that could lead to the learning of increasingly sophisticated mental abilities in a biologically plausible manner.”

Ullman's hand detector operates via a simple local optical flow calculation to detect 'mover' events.

“Ullman's hand detector, in fact, operates via a simple local optical flow calculation to detect “mover” events.”

• cites Ullman et al., 2012 sourceNot checked yet

  Not checked in the free run.

Optimization of initially unstructured 'blank slate' networks is not sufficient to generate complex cognition in the brain, even given a diversity of powerful genetically-specified cost functions and local learning rules.

“Optimization of initially unstructured “blank slate” networks is not sufficient to generate complex cognition in the brain, we argue, even given a diversity of powerful genetically-specified cost functions and local learning rules, as we have posited above.”

In Ullman et al. (2012), appearance, body context, and mover motion are all used to bootstrap off of one another, creating a detector that is better than any of its training heuristics.

“In Ullman et al. (2012), appearance, body context, and mover motion are all used to bootstrap off of one another, creating a detector that is better than any of its training heuristics.”

• cites Ullman et al., 2012 sourceNot checked yet

  Not checked in the free run.

An internally supervised learning algorithm containing specialized, hard-coded biases to detect hands, on the basis of their typical motion properties, can be used to bootstrap the training of an image recognition module that learns to recognize hands based on their appearance.

“Indeed, an internally supervised learning algorithm containing specialized, hard-coded biases to detect hands, on the basis of their typical motion properties, can be used to bootstrap the training of an image recognition module that learns to recognize hands based on their appearance.”

• cites Ullman et al., 2012 sourceNot checked yet

  Not checked in the free run.

A simple, hard-coded module bootstraps the training of a much more complex algorithm for visual recognition of hands.

“Thus, a simple, hard-coded module bootstraps the training of a much more complex algorithm for visual recognition of hands.”

The basic mechanisms of reinforcement learning may be widely re-purposed to train local networks using a variety of internally generated error signals.

“But beyond these well-known global reward signals, we argue that the basic mechanisms of reinforcement learning may be widely re-purposed to train local networks using a variety of internally generated error signals.”

The brain can use innate detectors to create cost functions and training procedures to train the next stages of learning.

“The brain can thus use innate detectors to create cost functions and training procedures to train the next stages of learning.”

Learning is a multi-phase process in which navigational cues are memorized prior to the acquisition of motor control.

“Learning is therefore a multi-phase process in which navigational cues are memorized prior to the acquisition of motor control.”

The brain may need specialized systems for storing and routing data, and for flexibly routing training signals such as target patterns, training data, reinforcement signals, attention signals, and modulatory signals.

“To allow this, the brain may need specialized systems for storing and routing data, and for flexibly routing training signals such as target patterns, training data, reinforcement signals, attention signals, and modulatory signals.”

Young children make inferences about causality selectively in situations where a human is trying to achieve an outcome.

“For example, young children make inferences about causality selectively in situations where a human is trying to achieve an outcome (Meltzoff et al., 2012, 2013).”

• cites Meltzoff et al., 2012, 2013Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Even within the cortex, microcircuitry within different areas may be specialized, allowing different cortical areas to adopt qualitatively distinct computational and learning configurations while sharing a common gross physical layout and communication interface.

“Even within the cortex itself, microcircuitry within different areas may be specialized: tinkered variations on a common ancestral microcircuit scaffold could potentially allow different cortical areas, such as sensory areas vs. prefrontal areas, to be configured to adopt a number of qualitatively distinct computational and learning configurations (Yuste et al., 2005; Marcus et al., 2014a,b), even”

• cites Yuste et al., 2005 sourceNot checked yet

  Not checked in the free run.

• cites Marcus et al., 2014a,bNot checked yet

  Not checked in the free run.

Some evidence suggests that the brain can rapidly detect faces without even a single feed-forward pass through the ventral visual stream.

“and indeed some evidence suggests that the brain can rapidly detect faces without even a single feed-forward pass through the ventral visual stream (Crouzet and Thorpe, 2011).”

• cites Crouzet and Thorpe, 2011Not checked yet

  Not checked in the free run.

These specialized structures evolutionarily pre-date the cortex.

“These structures evolutionarily pre-date (Lee et al., 2015) the cortex, and hence the cortex may have evolved to work in the context of such specialized mechanisms.”

• cites Lee et al., 2015 sourceNot checked yet

  Not checked in the free run.

Over forty distinct cell types have been found within cortex, differing in dendritic organization, distribution throughout the six cortical layers, connectivity pattern, gene expression, and electrophysiological properties.

“Within cortex, over forty distinct cell types—differing in such aspects as dendritic organization, distribution throughout the six cortical layers, connectivity pattern, gene expression, and electrophysiological properties—have already been found (Markram et al., 2015; Zeisel et al., 2015).”

• cites Markram et al., 2015 sourceNot checked yet

  Not checked in the free run.

• cites Zeisel et al., 2015 sourceNot checked yet

  Not checked in the free run.

The brain contains highly conserved structures, e.g., cortex, where it is theorized that a similar type of learning and/or computation is happening in multiple places.

“Looking at the brain, we indeed seem to find highly conserved structures, e.g., cortex, where it is theorized that a similar type of learning and/or computation is happening in multiple places (Braitenberg and Schutz, 1991; Douglas and Martin, 2004).”

• cites Braitenberg and Schutz, 1991Not checked yet

  Not checked in the free run.

• cites Douglas and Martin, 2004 sourceNot checked yet

  Not checked in the free run.

The brain seems to embody prior knowledge about the structure of emotional expressions and how they relate to causes in the world.

“Humans have areas devoted to emotional processing, and the brain seems to embody prior knowledge about the structure of emotional expressions and how they relate to causes in the world: emotions should have specific types of strong couplings to various other higher-level variables such as goal-satisfaction, should be expressed through the face, and so on (Phillips et al., 2002; Skerry and Spelke,”

• cites Phillips et al., 2002 sourceNot checked yet

  Not checked in the free run.

• cites Skerry and Spelke, 2014Not checked yet

  Not checked in the free run.

• cites Baillargeon et al., 2016 sourceNot checked yet

  Not checked in the free run.

Generalizing previous ideas, the brain uses optimization with respect to internally generated heuristic detection signals to bootstrap learning of biologically relevant features which would otherwise be missed by an unsupervised learner.

“Thus, generalizing previous ideas (Ullman et al., 2012; Poggio, 2015), we suggest that the brain uses optimization with respect to internally generated heuristic39 detection signals to bootstrap learning of biologically relevant features which would otherwise be missed by an unsupervised learner.”

• cites Ullman et al., 2012 sourceNot checked yet

  Not checked in the free run.

• cites Poggio, 2015Not checked yet

  Not checked in the free run.

The brain contains a large number of specialized structures, including thalamus, hippocampus, basal ganglia and cerebellum.

“But we also see a large number of specialized structures, including thalamus, hippocampus, basal ganglia and cerebellum (Solari and Stoner, 2011).”

• cites Solari and Stoner, 2011Not checked yet

  Not checked in the free run.

The cortex may have evolved to work in the context of such specialized mechanisms.

“These structures evolutionarily pre-date (Lee et al., 2015) the cortex, and hence the cortex may have evolved to work in the context of such specialized mechanisms.”

Special internally generated signals are needed for learning problems where standard unsupervised methods fail to discover properties of the world that are statistically weak but significant to the organism.

“Special, internally-generated signals are needed specifically for learning problems where standard unsupervised methods—based purely on matching the statistics of the world, or on optimizing simple mathematical objectives like temporal continuity or sparsity—will fail to discover properties of the world which are statistically weak in an objective sense but nevertheless have special significance”

• cites Ullman et al., 2012 sourceNot checked yet

  Not checked in the free run.

Innate detectors are used not as supervision signals per se, but rather to guide or steer the learning process, enabling it to discover features that would otherwise be difficult.

“In effect, the innate detectors are used not as supervision signals per se, but rather to guide or steer the learning process, enabling it to discover features that would otherwise be difficult.”

The training of optimization modules may need to be coordinated in a complex and dynamic fashion, including delivering the right training signals and activating the right learning rules in the right places and at the right times.

“Second, the training of optimization modules may need to be coordinated in a complex and dynamic fashion, including delivering the right training signals and activating the right learning rules in the right places and at the right times.”

Internally generated signals may allow a learning system to go beyond standard unsupervised methods by guiding or steering it to learn specific features or computations.

“These internally generated signals may allow a learning system to go beyond what can be learned via standard unsupervised methods, effectively guiding or steering the system to learn specific features or computations (Ullman et al., 2012).”

• cites Ullman et al., 2012 sourceNot checked yet

  Not checked in the free run.

The Scrub-Jay can remember the place and time of last visit for hundreds of different locations, e.g., to determine whether high-quality food is currently buried at any given location.

“For example, the Scrub-Jay can remember the place and time of last visit for hundreds of different locations, e.g., to determine whether high-quality food is currently buried at any given location (Clayton and Dickinson, 1998).”

• cites Clayton and Dickinson, 1998 sourceNot checked yet

  Not checked in the free run.

The hippocampus appears to explore, in simulation, possible future trajectories to a goal, even those involving previously unvisited locations.

“The hippocampus appears to explore, in simulation, possible future trajectories to a goal, even those involving previously unvisited locations (Ólafsdóttir et al., 2015).”

• cites Ólafsdóttir et al., 2015 sourceNot checked yet

  Not checked in the free run.

Progress is being made in both neuroscience and machine learning on finding potential mechanisms for hierarchical planning and goal-seeking.

“Progress is being made in both neuroscience and machine learning on finding potential mechanisms for this type of hierarchical planning and goal-seeking.”

• cites O'Reilly et al., 2014aNot checked yet

  Not checked in the free run.

• cites Huys et al., 2015 sourceNot checked yet

  Not checked in the free run.

• cites Balaguer et al., 2016 sourceNot checked yet

  Not checked in the free run.

There is preliminary evidence that social relationships may be represented through a hippocampal map.

“as well as preliminary evidence that social relationships may also be represented through a hippocampal map (Tavares et al., 2015).”

• cites Tavares et al., 2015 sourceNot checked yet

  Not checked in the free run.

Human language syntax has a hierarchical structure, described as composition of smaller forms like words and phrases into larger ones.

“Human language syntax has a hierarchical structure, which Berwick et al described as “composition of smaller forms like words and phrases into larger ones” (Berwick et al., 2012; Miyagawa et al., 2013).”

• cites Berwick et al., 2012 sourceNot checked yet

  Not checked in the free run.

• cites Miyagawa et al., 2013 sourceNot checked yet

  Not checked in the free run.

There is evidence that the brain uses a hierarchical strategy for control problems.

“There is evidence that the brain uses such a strategy (Botvinick et al., 2009; Botvinick and Weinstein, 2014), beside neural network demonstrations (Wayne and Abbott, 2014).”

• cites Botvinick et al., 2009 sourceNot checked yet

  Not checked in the free run.

• cites Botvinick and Weinstein, 2014Not checked yet

  Not checked in the free run.

• cites Wayne and Abbott, 2014Not checked yet

  Not checked in the free run.

This type of architecture can simplify the learning of control problems.

“Thus this type of architecture can simplify the learning of control problems.”

Sequential programs enable one-shot learning of new geometric/visual concepts.

“In particular, sequential programs were found to enable one-shot learning of new geometric/visual concepts (Lake et al., 2015).”

• cites Lake et al., 2015 sourceNot checked yet

  Not checked in the free run.

In a hierarchical factorization of motor control, every layer has a characteristic timescale.

“and second, every layer has a characteristic timescale.”

Humans excel at stitching together sub-actions to form larger actions.

“Humans excel at stitching together sub-actions to form larger actions (Verwey, 1996; Acuna et al., 2014; Sejnowski and Poizner, 2014).”

• cites Verwey, 1996Not checked yet

  Not checked in the free run.

• cites Acuna et al., 2014 sourceNot checked yet

  Not checked in the free run.

• cites Sejnowski and Poizner, 2014Not checked yet

  Not checked in the free run.

In some simple models, targeted navigation in the hippocampus is achieved via the dynamics of bump attractors or propagating waves in a place cell network with Hebbian plasticity and adaptation.

“In some simple models, targeted navigation in the hippocampus is achieved via the dynamics of “bump attractors” or propagating waves in a place cell network with Hebbian plasticity and adaptation (Hopfield, 2009; Buzsáki and Moser, 2013; Ponulak and Hopfield, 2013)”

• cites Hopfield, 2009 sourceNot checked yet

  Not checked in the free run.

• cites Buzsáki and Moser, 2013 sourceNot checked yet

  Not checked in the free run.

• cites Ponulak and Hopfield, 2013Not checked yet

  Not checked in the free run.

Broca's area appears to be specifically involved in processing the hierarchical structure of behavior, as such, as opposed to its detailed temporal structure.

“Some brain areas like Broca's area, known for its involvement in language, also appear to be specifically involved in processing the hierarchical structure of behavior, as such, as opposed to its detailed temporal structure (Koechlin and Jubault, 2006).”

• cites Koechlin and Jubault, 2006Not checked yet

  Not checked in the free run.

In animals, the hippocampus aids-at least in part through place cell and grid cell systems-in efficient learning about new environments and in targeted navigation in such environments.

“It appears that, in animals, the hippocampus aids—at least in part through “place cell” and “grid cell” systems—in efficient learning about new environments and in targeted navigation in such environments (Brown et al., 2016).”

• cites Brown et al., 2016 sourceNot checked yet

  Not checked in the free run.

The spinal cord suggests the existence of two levels of a hierarchy, each trained using different cost functions.

“So the spinal cord itself already suggests the existence of two levels of a hierarchy, each trained using different cost functions.”

Central processing in the brain has an effectively simplified physical system to control, e.g., one that is far more linear.

“Nevertheless, the effect is that central processing in the brain has an effectively simplified physical system to control, e.g., one that is far more linear.”

There is preliminary evidence from fMRI that abstract concepts are represented according to grid-cell-like hexagonal grid structures in humans.

“Interestingly, there is preliminary evidence from fMRI that abstract concepts are also represented according to grid-cell-like hexagonal grid structures in humans (Constantinescu et al., 2016)”

• cites Constantinescu et al., 2016 sourceNot checked yet

  Not checked in the free run.

Model-based hierarchical reinforcement learning and prospective optimization implicates a network coupling the dorsolateral prefrontal and orbitofrontal cortex, and the ventral and dorsolateral striatum.

“The study of model-based hierarchical reinforcement learning and prospective optimization (Sejnowski and Poizner, 2014), which concerns the planning and evaluation of nested sequences of actions, implicates a network coupling the dorsolateral prefontral and orbitofrontal cortex, and the ventral and dorsolateral striatum (Botvinick et al., 2009).”

• cites Sejnowski and Poizner, 2014Not checked yet

  Not checked in the free run.

• cites Botvinick et al., 2009 sourceNot checked yet

  Not checked in the free run.

Once an environment becomes familiar, areas of the neocortex can take over the role of navigation.

“Interestingly, once an environment becomes familiar, it appears that areas of the neocortex can take over the role of navigation (Hasselmo and Stern, 2015).”

• cites Hasselmo and Stern, 2015Not checked yet

  Not checked in the free run.

Goal hierarchies are beginning to be incorporated into deep learning.

“Goal hierarchies are beginning to be incorporated into deep learning (Kulkarni et al., 2016).”

• cites Kulkarni et al., 2016Not checked yet

  Not checked in the free run.

Structured, serial, hierarchical probabilistic programs have recently been shown to model aspects of human conceptual representation and compositional learning.

“Structured, serial, hierarchical probabilistic programs have recently been shown to model aspects of human conceptual representation and compositional learning (Lake et al., 2015).”

• cites Lake et al., 2015 sourceNot checked yet

  Not checked in the free run.

Human reward systems reflect changing goals and subgoals.

“At the highest level of the decision making and control hierarchy, human reward systems reflect changing goals and subgoals”

• cites Buschman and Miller, 2014Not checked yet

  Not checked in the free run.

• cites O'Reilly et al., 2014bNot checked yet

  Not checked in the free run.

• cites Pezzulo et al., 2014 sourceNot checked yet

  Not checked in the free run.

Converting the computationally hard optimal control problem into a hierarchical approximation promises to make it dramatically easier.

“Converting the computationally hard optimal control problem into a hierarchical approximation promises to make it dramatically easier.”

In a hierarchical factorization of motor control, each level can be solved with its own cost functions.

“First, each level can be solved with its own cost functions”

Higher-level cognitive tasks such as prospective planning appear to share computational sub-problems with path-finding.

“Higher-level cognitive tasks such as prospective planning appear to share computational sub-problems with path-finding (Hassabis and Maguire, 2009)52.”

• cites Hassabis and Maguire, 2009 sourceNot checked yet

  Not checked in the free run.

Goal representations emerge in the amygdala, prefrontal cortex and other areas.

“with goal representations themselves emerging in the amygdala, prefrontal cortex and other areas (O'Reilly et al., 2014a).”

• cites O'Reilly et al., 2014aNot checked yet

  Not checked in the free run.

Many of the control problems we appear to be solving are hierarchical.

“Importantly, many of the control problems we appear to be solving are hierarchical.”

Multiple different kinds of memory are needed.

“Importantly, multiple different kinds of memory are needed (Squire, 2004).”

• cites Squire, 2004 sourceNot checked yet

  Not checked in the free run.

Visual data streams can be segmented into elements such as foreground versus background and objects that can move of their own accord versus those that cannot, using diverse unsupervised criteria.

“For example, visual data streams can be segmented into elements like foreground vs. background, objects that can move of their own accord vs. those that cannot, all using diverse unsupervised criteria (Ullman et al., 2012; Poggio, 2015).”

• cites Ullman et al., 2012 sourceNot checked yet

  Not checked in the free run.

• cites Poggio, 2015Not checked yet

  Not checked in the free run.

These detailed pathways underlying the construction of cost functions for vocal learning are beginning to be mapped.

“These detailed pathways underlying the construction of cost functions for vocal learning are beginning to be mapped (Mandelblat-Cerf et al., 2014).”

• cites Mandelblat-Cerf et al., 2014 sourceNot checked yet

  Not checked in the free run.

Various techniques exist to carry out comparisons of optimized receptive fields with measured receptive fields in fMRI studies, including population receptive field estimation and representational dissimilarity matrices.

“Various techniques exist to carry out such comparisons in fRMI studies, including population receptive field estimation (Dumoulin and Wandell, 2008; Güçlü and van Gerven, 2015) and representational dissimilarity matrices (Kriegeskorte et al., 2008; Khaligh-Razavi and Kriegeskorte, 2014).”

• cites Dumoulin and Wandell, 2008Not checked yet

  Not checked in the free run.

• cites Güçlü and van Gerven, 2015 sourceNot checked yet

  Not checked in the free run.

• cites Kriegeskorte et al., 2008 sourceNot checked yet

  Not checked in the free run.

The brain is not solely an end-to-end reinforcement trained system.

“we argue that, though important, they are only a part of the picture, i.e., that the brain is not solely an end-to-end reinforcement trained system.”

The role of the hippocampus in imagination, concept generation, scene construction, mental exploration and goal-directed path planning suggests that it could help to create generative models to underpin more complex inference such as program induction or common-sense world simulation.

“The role of the hippocampus in imagination, concept generation (Kumaran et al., 2009), scene construction (Hassabis and Maguire, 2007), mental exploration and goal-directed path planning (Hopfield, 2009; Ólafsdóttir et al., 2015; Brown et al., 2016) suggests that it could help to create generative models to underpin more complex inference such as program induction (Lake et al., 2015) or common-s”

• cites Kumaran et al., 2009 sourceNot checked yet

  Not checked in the free run.

• cites Hassabis and Maguire, 2007 sourceNot checked yet

  Not checked in the free run.

• cites Hopfield, 2009 sourceNot checked yet

  Not checked in the free run.

It was recently shown that complex probabilistic population coding and inference can arise automatically from backpropagation-based training of simple neural networks, without needing to be built in by hand.

“For example, it was recently shown that complex probabilistic population coding and inference can arise automatically from backpropagation-based training of simple neural networks (Orhan and Ma, 2016), without needing to be built in by hand.”

• cites Orhan and Ma, 2016Not checked yet

  Not checked in the free run.

The error signals that may be used to train specific sub-networks in the brain, via some approximation of gradient descent or otherwise, are not necessarily equivalent to reward signals.

“The error signals that may be used to train specific sub-networks in the brain, via some approximation of gradient descent or otherwise, are not necessarily equivalent to reward signals.”

This has already been done to aid the interpretation of the neural dynamics underlying decision making in the PFC, working memory in the posterior parietal cortex, and object or action representation in the visual system.

“This has already been done to aid the interpretation of the neural dynamics underlying decision making in the PFC (Sussillo, 2014), working memory in the posterior parietal cortex (Rajan et al., 2016) and object or action representation in the visual system (Tacchetti et al., 2016; Yamins and DiCarlo, 2016a,b).”

• cites Sussillo, 2014 sourceNot checked yet

  Not checked in the free run.

• cites Rajan et al., 2016 sourceNot checked yet

  Not checked in the free run.

• cites Tacchetti et al., 2016Not checked yet

  Not checked in the free run.

In computer vision, researchers have begun to re-appropriate neural networks trained for one task and then deploy them on new tasks other than the ones they were trained for or for which more limited training data is available.

“In computer vision, we have begun to see researchers re-appropriate neural networks trained for one task (e.g., ImageNet classification) and then deploy them on new tasks other than the ones they were trained for or for which more limited training data is available (Oquab et al., 2014; Yosinski et al., 2014; Noroozi and Favaro, 2016).”

• cites Oquab et al., 2014Not checked yet

  Not checked in the free run.

• cites Yosinski et al., 2014Not checked yet

  Not checked in the free run.

• cites Noroozi and Favaro, 2016 sourceNot checked yet

  Not checked in the free run.

Many of the most notable successes of deep learning, from language modeling to vision to motor control, have been driven by end-to-end optimization of single task objectives.

“Many of the most notable successes of deep learning, from language modeling (Sutskever et al., 2011), to vision (Krizhevsky et al., 2012), to motor control (Levine et al., 2015), have been driven by end-to-end optimization of single task objectives.”

• cites Sutskever et al., 2011Not checked yet

  Not checked in the free run.

• cites Krizhevsky et al., 2012Not checked yet

  Not checked in the free run.

• cites Levine et al., 2015Not checked yet

  Not checked in the free run.

The success of deep learning is driven by the separation of learning into an optimization algorithm (backpropagation) and approaches to design a cost function and training procedure that yield the desired computation.

“Driving the success is the separation of the problem of learning into two pieces: (1) An algorithm, backpropagation, that allows efficient distributed optimization, and (2) Approaches to turn any given problem into an optimization problem, by designing a cost function and training procedure which will result in the desired computation.”

The search for biologically plausible forms of backpropagation has already led to insights such as random feedback weights (feedback alignment) and the unexpected power of internal FORCE learning in chaotic, spontaneously active recurrent networks.

“Moreover, the search for biologically plausible forms of backpropagation has already led to interesting insights, such as the possibility of using random feedback weights (feedback alignment) in backpropagation (Lillicrap et al., 2014), or the unexpected power of internal FORCE learning in chaotic, spontaneously active recurrent networks (Sussillo and Abbott, 2009).”

• cites Lillicrap et al., 2014Not checked yet

  Not checked in the free run.

• cites Sussillo and Abbott, 2009 sourceNot checked yet

  Not checked in the free run.

This program is beginning to be carried out in the fly brain, in which twenty specific types of dopamine neuron project to distinct anatomical compartments of the mushroom body to train distinct odor classifiers operating on a set of high-dimensional odor representations.

“This program is beginning to be carried out in the fly brain, in which twenty specific types of dopamine neuron project to distinct anatomical compartments of the mushroom body to train distinct odor classifiers operating on a set of high-dimensional odor representations (Caron et al., 2013; Aso et al., 2014a,b; Cohn et al., 2015).”

• cites Caron et al., 2013 sourceNot checked yet

  Not checked in the free run.

• cites Aso et al., 2014a,bNot checked yet

  Not checked in the free run.

• cites Cohn et al., 2015 sourceNot checked yet

  Not checked in the free run.

Due to the complexity and variability of the brain, pure bottom-up analysis of neural data faces potential challenges of interpretation.

“Due to the complexity and variability of the brain, pure “bottom up” analysis of neural data faces potential challenges of interpretation (Robinson, 1992; Jonas and Kording, 2016).”

• cites Robinson, 1992 sourceNot checked yet

  Not checked in the free run.

• cites Jonas and Kording, 2016 sourceNot checked yet

  Not checked in the free run.

Many top-down frameworks for understanding neural computation boil down to assuming that the brain simply optimizes a single, given cost function for a single computational architecture.

“Interestingly, many of the “top down” frameworks boil down to assuming that the brain simply optimizes a single, given cost function for a single computational architecture.”

• cites Pinker, 1999 sourceNot checked yet

  Not checked in the free run.

• cites Marcus, 2001 sourceNot checked yet

  Not checked in the free run.

• cites Bialek, 2002 sourceNot checked yet

  Not checked in the free run.