CoChat

What Caused What? A Quantitative Account of Actual Causation Using Dynamical Causal Networks

Open access
Albantakis, Larissa, Marshall, William, Hoel, Erik, Tononi, Giulio · 2019 · Entropy

The DOI matches its Crossref record; 65 of the 66 references it cites were found in a registry or an index, and 1 are not checked yet.

Abstract

Actual causation is concerned with the question: “What caused what?” Consider a transition between two states within a system of interacting elements, such as an artificial neural network, or a biological brain circuit. Which combination of synapses caused the neuron to fire? Which image features caused the classifier to misinterpret the picture? Even detailed knowledge of the system’s causal network, its elements, their states, connectivity, and dynamics does not automatically provide a straightforward answer to the “what caused what?” question. Counterfactual accounts of actual causation, based on graphical models paired with system interventions, have demonstrated initial success in addressing specific problem cases, in line with intuitive causal judgments. Here, we start from a set of basic requirements for causation (realization, composition, information, integration, and exclusion) and develop a rigorous, quantitative account of actual causation, that is generally applicable to discrete dynamical systems. We present a formal framework to evaluate these causal requirements based on system interventions and partitions, which considers all counterfactuals of a state transition. This framework is used to provide a complete causal account of the transition by identifying and quantifying the strength of all actual causes and effects linking the two consecutive system states. Finally, we examine several exemplary cases and paradoxes of causation and show that they can be illuminated by the proposed framework for quantifying actual causation.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.3390/e21050459

Cites

65 of 66 references verified0 citation errors · 0 not checkable

Statistics

Figures

Figure A1 Reducible and irreducible causal accounts. (A) “Prevention” example (see Figure 7D, main text). We have that A=0 bits as {A=1} does not contribute to any causal links. (B) Irreducible transition (see Figure 6, main text). A partition of the transition along the MIP destroys the 2nd-order causal link, leading to A=0.17 bits. (C) In larger systems, A can be used to identify (sub-)transitions with highly entangled causes and effect. While the causes and effects in the full transition are only weakly entangled, with A=0.03 bits, the top and bottom (sub-)transitions are irreducible, with A=0.83 bits.
Figure A1 Reducible and irreducible causal accounts. (A) “Prevention” example (see Figure 7D, main text). We have that A=0 bits as {A=1} does not contribute to any causal links. (B) Irreducible transition (see Figure 6, main text). A partition of the transition along the MIP destroys the 2nd-order causal link, leading to A=0.17 bits. (C) In larger systems, A can be used to identify (sub-)transitions with highly entangled causes and effect. While the causes and effects in the full transition are only weakly entangled, with A=0.03 bits, the top and bottom (sub-)transitions are irreducible, with A=0.83 bits.CC BY · PubMed Central
Figure 1 Realization: Dynamical causal network and transition. (A) A discrete dynamical system constituted of two interacting elements: An OR- and AND-logic gate, which are updated synchronously at every time step, according to their input-output functions. Arrows denote connections between the elements. (B) The same system can be represented as a dynamical causal network over consecutive time steps. (C) The system described by its entire set of transition probabilities. As this particular system is deterministic, all transitions have a probability of either p=0 or p=1. (D) A realization of a system transient over two time steps, consistent with the system’s transition probabilities: {(OR,AND)t−1=10}≺{(OR,AND)t=10}.
Figure 1 Realization: Dynamical causal network and transition. (A) A discrete dynamical system constituted of two interacting elements: An OR- and AND-logic gate, which are updated synchronously at every time step, according to their input-output functions. Arrows denote connections between the elements. (B) The same system can be represented as a dynamical causal network over consecutive time steps. (C) The system described by its entire set of transition probabilities. As this particular system is deterministic, all transitions have a probability of either p=0 or p=1. (D) A realization of a system transient over two time steps, consistent with the system’s transition probabilities: {(OR,AND)t−1=10}≺{(OR,AND)t=10}.CC BY · PubMed Central
Figure 2 Assessing cause and effect repertoires. (A) Example effect repertoires, indicating how the occurrence {ORt−1=1} constrains the future states of ORt (left) and (OR,AND)t (right) in the causal network shown in Figure 1. (B) Example cause repertoires indicating how the occurrences {ORt=1} (left) and {(OR,AND)t=10} (right) constrain the past states of ORt−1. Throughout the manuscript, filled circles denote occurrences, while open circles denote candidate causes and effects. Green shading is used for t, blue for t−1. Nodes that are not included in the occurrence or candidate cause/effect are causally marginalized.
Figure 2 Assessing cause and effect repertoires. (A) Example effect repertoires, indicating how the occurrence {ORt−1=1} constrains the future states of ORt (left) and (OR,AND)t (right) in the causal network shown in Figure 1. (B) Example cause repertoires indicating how the occurrences {ORt=1} (left) and {(OR,AND)t=10} (right) constrain the past states of ORt−1. Throughout the manuscript, filled circles denote occurrences, while open circles denote candidate causes and effects. Green shading is used for t, blue for t−1. Nodes that are not included in the occurrence or candidate cause/effect are causally marginalized.CC BY · PubMed Central
Figure 3 Partitioning the repertoire π(Yt∣xt−1). (A) The set of all possible partitions of an occurrence, Ψ(xt−1,Yt), includes all partitions of xt−1 into 2≤m≤|xt−1| parts, according to Equation (7); as well as the special case ψ={(xt−1,⌀)}. Considering this special case a potential partition has the added benefit of allowing us to treat singleton occurrences and multi-variate occurrences in a common framework. (B) Except for the special case when the occurrence is completely cut from the nodes it constrains, we generally do not consider cases with m=1 as partitions of the occurrence. The partition must eliminate the possibility of joint constraints of xt−1 onto Yt. The set of all partitions Ψ(Xt−1,yt) of a cause repertoire π(Xt−1∣yt) includes all partitions of yt into 2≤m≤|yt| parts, according to Equation (9), and, again, the special case of ψ={(⌀,yt)} for m=1.
Figure 3 Partitioning the repertoire π(Yt∣xt−1). (A) The set of all possible partitions of an occurrence, Ψ(xt−1,Yt), includes all partitions of xt−1 into 2≤m≤|xt−1| parts, according to Equation (7); as well as the special case ψ={(xt−1,⌀)}. Considering this special case a potential partition has the added benefit of allowing us to treat singleton occurrences and multi-variate occurrences in a common framework. (B) Except for the special case when the occurrence is completely cut from the nodes it constrains, we generally do not consider cases with m=1 as partitions of the occurrence. The partition must eliminate the possibility of joint constraints of xt−1 onto Yt. The set of all partitions Ψ(Xt−1,yt) of a cause repertoire π(Xt−1∣yt) includes all partitions of yt into 2≤m≤|yt| parts, according to Equation (9), and, again, the special case of ψ={(⌀,yt)} for m=1.CC BY · PubMed Central
Figure 4 Considering the power set of occurrences. All subsets xt−1⊆vt−1 and yt⊆vt are considered as occurrences which may have an actual effect or an actual cause.
Figure 4 Considering the power set of occurrences. All subsets xt−1⊆vt−1 and yt⊆vt are considered as occurrences which may have an actual effect or an actual cause.CC BY · PubMed Central
Figure 5 Assessing the cause and effect information, their irreducibility (integration), and the maximum cause/effect (exclusion). (A,B) Example effect and cause information. The state that actually occurred is selected from the effect or cause repertoire (green is used for effects, blue for causes). Its probability is compared to the probability of the same state when unconstrained (overlaid distributions without fill). All repertoires are based on product probabilities, π (Equations (3) and (4)), that discount correlations due to common inputs when variables are causally marginalized. For example, π({(OR,AND)t=01})>0 in (A, right panel), although p({(OR,AND)t=01})=0. (C,D) Integrated effect and cause information. The probability of the actual state in the effect or cause repertoire is compared against its probability in the partitioned effect or cause repertoire (overlaid distributions without fill). Of all second-order occurrences shown, only {(OR,AND)t=10} irreducibly constrains {(OR,AND)t−1=10}. For first-order occurrences, αc/e=ρc/e (see text). Maximum values are highlighted in bold. If, as in panel (B), a superset of a candidate cause or effect specifies the same maximum value, it is excluded by a minimality condition.
Figure 5 Assessing the cause and effect information, their irreducibility (integration), and the maximum cause/effect (exclusion). (A,B) Example effect and cause information. The state that actually occurred is selected from the effect or cause repertoire (green is used for effects, blue for causes). Its probability is compared to the probability of the same state when unconstrained (overlaid distributions without fill). All repertoires are based on product probabilities, π (Equations (3) and (4)), that discount correlations due to common inputs when variables are causally marginalized. For example, π({(OR,AND)t=01})>0 in (A, right panel), although p({(OR,AND)t=01})=0. (C,D) Integrated effect and cause information. The probability of the actual state in the effect or cause repertoire is compared against its probability in the partitioned effect or cause repertoire (overlaid distributions without fill). Of all second-order occurrences shown, only {(OR,AND)t=10} irreducibly constrains {(OR,AND)t−1=10}. For first-order occurrences, αc/e=ρc/e (see text). Maximum values are highlighted in bold. If, as in panel (B), a superset of a candidate cause or effect specifies the same maximum value, it is excluded by a minimality condition.CC BY · PubMed Central
Figure 6 Causal Account. There are two first-order occurrences with actual effects and actual causes. In addition, the second-order occurrence {(OR,AND)t=10} has an actual cause {(OR,AND)t−1=10}.
Figure 6 Causal Account. There are two first-order occurrences with actual effects and actual causes. In addition, the second-order occurrence {(OR,AND)t=10} has an actual cause {(OR,AND)t−1=10}.CC BY · PubMed Central
Figure 7 Four dynamically identical transitions can have different causal accounts. Shown are the transitions (top) and their respective causal accounts (bottom).
Figure 7 Four dynamically identical transitions can have different causal accounts. Shown are the transitions (top) and their respective causal accounts (bottom).CC BY · PubMed Central
Figure 8 A linear threshold unit with four inputs and threshold k=3 (Majority gate). (A) All inputs are considered relevant variables. (B) The case D=0 is taken as a fixed background condition (indicated by the red pin).
Figure 8 A linear threshold unit with four inputs and threshold k=3 (Majority gate). (A) All inputs are considered relevant variables. (B) The case D=0 is taken as a fixed background condition (indicated by the red pin).CC BY · PubMed Central
Figure 9 Disjunction of two conjunctions (A∧B)∨C. (A) All inputs to D are considered relevant variables. (B) B=0 is taken as a fixed background condition.
Figure 9 Disjunction of two conjunctions (A∧B)∨C. (A) All inputs to D are considered relevant variables. (B) B=0 is taken as a fixed background condition.CC BY · PubMed Central
Figure 10 Complicated voting.
Figure 10 Complicated voting.CC BY · PubMed Central
Figure 11 Voting with three possible candidates.
Figure 11 Voting with three possible candidates.CC BY · PubMed Central
Figure 12 Probabilistic variables. While the transition shown in (A) does have a deterministic equivalent, the transition shown in (B) would be impossible in the deterministic case.
Figure 12 Probabilistic variables. While the transition shown in (A) does have a deterministic equivalent, the transition shown in (B) would be impossible in the deterministic case.CC BY · PubMed Central
Figure 13 Simple classifier. D is a “dot-detector”, S is a “segment-detector”, and L is a “line-detector” (see text).
Figure 13 Simple classifier. D is a “dot-detector”, S is a “segment-detector”, and L is a “line-detector” (see text).CC BY · PubMed Central
Figure 14 Causal conditioning and marginalizing. (A) Variables outside the transition of interest are treated as fixed background conditions (indicated by the red pins). The transition probabilities p(vt|vt−1) are conditioned on the state of these background variables. (B) When evaluating the strength of a causal link within the transition, the remaining variables in Gu outside the causal link are causally marginalized; that is, they are replaced by an average across all their possible states. With B marginalized, the state of A by itself does not determine and is not determined by the occurrence {XOR=1}.
Figure 14 Causal conditioning and marginalizing. (A) Variables outside the transition of interest are treated as fixed background conditions (indicated by the red pins). The transition probabilities p(vt|vt−1) are conditioned on the state of these background variables. (B) When evaluating the strength of a causal link within the transition, the remaining variables in Gu outside the causal link are causally marginalized; that is, they are replaced by an average across all their possible states. With B marginalized, the state of A by itself does not determine and is not determined by the occurrence {XOR=1}.CC BY · PubMed Central
Figure 15 Composition: High-order occurrences. (A) Double Bi-conditional: Transition and causal account. (B) Cause repertoires corresponding to the two first-order and one second-order occurrences with actual causes (see text).
Figure 15 Composition: High-order occurrences. (A) Double Bi-conditional: Transition and causal account. (B) Cause repertoires corresponding to the two first-order and one second-order occurrences with actual causes (see text).CC BY · PubMed Central
Figure 16 Integration: Irreducible versus reducible occurrences. (A) Transition and causal account of Figure 6. (B) The second-order occurrence {(OR,AND)=10} with actual cause {AB=10} is irreducible under the MIP. (C) Reducible transition with equivalent first-order causal links, but missing the second-order causal link present in (A). (D) The constraints specified by the second-order occurrence {(OR,AND)=10} here are the same, and thus reducible, to those under the MIP.
Figure 16 Integration: Irreducible versus reducible occurrences. (A) Transition and causal account of Figure 6. (B) The second-order occurrence {(OR,AND)=10} with actual cause {AB=10} is irreducible under the MIP. (C) Reducible transition with equivalent first-order causal links, but missing the second-order causal link present in (A). (D) The constraints specified by the second-order occurrence {(OR,AND)=10} here are the same, and thus reducible, to those under the MIP.CC BY · PubMed Central
Figure 17 Exclusion: Any occurrence can, at most, have one actual cause or effect. (A) Out of the three candidate causes {A=1}, {B=1}, and {AB=11}, the actual cause of {AND=1} is the high-order occurrence {AB=11}, with αc=αc^max=2.0 bits. (B) Out of the three candidate effects, {AND=1}, {XOR=1}, and {(AND,XOR)=11}, the actual effect of {A=1} is the first-order occurrence {AND=1}, with αe=αe^max=1.0 bit; {(AND,XOR)=11} is excluded by the minimality condition (Definition 2).
Figure 17 Exclusion: Any occurrence can, at most, have one actual cause or effect. (A) Out of the three candidate causes {A=1}, {B=1}, and {AB=11}, the actual cause of {AND=1} is the high-order occurrence {AB=11}, with αc=αc^max=2.0 bits. (B) Out of the three candidate effects, {AND=1}, {XOR=1}, and {(AND,XOR)=11}, the actual effect of {A=1} is the first-order occurrence {AND=1}, with αe=αe^max=1.0 bit; {(AND,XOR)=11} is excluded by the minimality condition (Definition 2).CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

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?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 66 references

Research works printed but not linked: 6

  1. Lewis, D. (1973). Counterfactuals, Harvard University Press.
  2. Rubenstein, P.K., Weichwald, S., Bongers, S., Mooij, J.M., Janzing, D., Grosse-Wentrup, M., and Schölkopf, B. (2017). Causal Consistency of Structural Equation Models. arXiv.
  3. Pearl, J. (2009). The Structural Theory of Causations. Causality in the Sciences, Oxford University Press. Number July.
  4. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., and Fergus, R. (2013). Intriguing properties of neural networks. arXiv.
  5. Knight, 2017, The dark art at the heart of AI, MIT Technol. Rev., 120, 55
  6. Damasio, A.R., and Damasio, H. (2012). Neurobiology of Decision-Making, Springer. Research and Perspectives in Neurosciences.

Cited by

Metadata

Authors
Albantakis, L., Marshall, W., Hoel, E., & Tononi, G.
Affiliations
  • University of Wisconsin–Madison (US)
  • Tufts University (US)
Corresponding author
  • Albantakis, Larissa
  • Tononi, Giulio
Venue
Entropy
Publisher
MDPI AG
Published
2019-05-02
Volume
21
Issue
5
Pages
459
ISSN
1099-4300 (ISSN-L)
OpenAlex
W2908619233
Cited by
67
References
64
Citations by year
2026: 4 · 2025: 7 · 2024: 2 · 2023: 9 · 2022: 7 · 2021: 12 · 2020: 15 · 2019: 8
FWCI
4.79
Type
journal article
Language
en
Open access
Locations
  • Entropy (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • arXiv (Cornell University)
Funders
Templeton World Charity Foundation (TWCF0067/AB41, TWCF0196)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Bayesian Modeling and Causal Inference
  • Social Sciences › Arts and Humanities › History and Philosophy of Science › Philosophy of Biology and Scientific Explanation
  • Physical Sciences › Computer Science › Artificial Intelligence › Logic, Reasoning, and Knowledge
Keywords
actual causation, counterfactual reasoning, systemic intervention, discrete dynamical systems, state transitions, probabilistic graphical models

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.

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

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.