Causal Learning with Neural Assemblies

arxiv: 2604.26919 · v1 · submitted 2026-04-29 · 💻 cs.LG · cs.AI· cs.NE

Causal Learning with Neural Assemblies

Evangelia Kopadi , Dimitris Kalles This is my paper

Pith reviewed 2026-05-07 10:35 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.NE

keywords neural assembliesdirectional learningcausal discoverylocal plasticitysynaptic asymmetrybiologically plausible learningDIRECT mechanism

0 comments p. Extension

The pith

Neural assemblies can learn causal directionality using only their built-in operations of projection, local plasticity control, and sparse winner selection.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that groups of neurons firing together can internalize which variable influences which other one, without needing global error signals or backpropagation. It introduces a mechanism called DIRECT that co-activates a source assembly and a target assembly under an adaptive gain schedule so that directed relations become embedded in the synaptic weights. Validation relies on two readouts: the resulting asymmetry in forward versus reverse synaptic strengths, and the reliability of signal flow in one direction only. If this holds, it supplies an auditable, biologically local route from neural dynamics to formal causal models, where each causal claim traces back to specific winning neurons and measurable weight differences.

Core claim

The inherent operations of neural assemblies—projection, local plasticity control, and sparse winner selection—are sufficient for directional learning. DIRECT co-activates source and target assemblies under an adaptive gain schedule to internalize directed relations, yielding perfect structural recovery under supervised known-structure conditions across domains. The resulting causal claims are auditable because they rest on synaptic-strength asymmetry and functional propagation overlap rather than opaque global optimization.

What carries the argument

DIRECT (DIRectional Edge Coupling/Training), a local co-activation rule that pairs source and target assemblies under an adaptive gain schedule to produce measurable synaptic asymmetry and unidirectional propagation.

If this is right

Causal claims become traceable to specific neural winners and measurable synaptic asymmetries rather than global loss surfaces.
The same local operations that support classification and planning can now support directional causal graphs.
Perfect recovery of known causal structures is achieved across multiple domains without backpropagation.
The framework supplies an explicit bridge between spike-based local dynamics and formal causal semantics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same local rule could be tested in neuromorphic hardware where global backpropagation is unavailable.
If the adaptive gain schedule can be made fully unsupervised, the method might extend to causal discovery from observational data alone.
Disrupting sparse winner selection while preserving co-activation should selectively abolish directional learning while leaving undirected associations intact.

Load-bearing premise

Co-activating source and target assemblies under an adaptive gain schedule is enough to embed directed relations that pass both synaptic-asymmetry and functional-propagation checks.

What would settle it

Run the co-activation schedule on a known directed edge and observe whether the forward synaptic weights remain statistically indistinguishable from the reverse weights or whether activity from the target assembly fails to propagate back to the source more reliably than the reverse.

Figures

Figures reproduced from arXiv: 2604.26919 by Dimitris Kalles, Evangelia Kopadi.

**Figure 1.** Figure 1: Pipeline overview. This investigation continues the same research trajectory developed in section 7. In the first phase, we established that neural assemblies can preserve causal information across a full end-to-end process, from raw variables to neuron and assembly feature spaces, under controlled conditions. Building on that foundation, the present phase asks the next logical question: can assemblies g… view at source ↗

**Figure 2.** Figure 2: Warm+ramp success diagnostic. The need for this adaptive schedule becomes apparent when compared with the baseline parallel setting, where formation and directional updates run concurrently without adaptive tuning. In that regime, winner drift remains high, directional signal becomes noisy, and edge-level asymmetry is harder to attribute mechanistically view at source ↗

**Figure 3.** Figure 3: Parallel failure diagnostic. Second, a single Brain simulator is shared across all stages, ensuring that formation and binding operate on the same connectome (the model’s full synaptic wiring matrix) and plasticity state. Third, each stage and the subsequent readout step are independently inspectable, so that if a directed link is not recovered, the discrepancy can be localized to assembly instability, … view at source ↗

**Figure 5.** Figure 5: Causal information preservation process. view at source ↗

**Figure 4.** Figure 4: Mixed-input encoding overview. Across domains, causal discovery on assembly features closely matched discovery on neuron features when using the PC algorithm (Peter-Clark conditional-independence method) [15], and largely recovered the ground-truth structure. Interventional checks using do-operator data generation (do(X=x); explicit intervention that sets a variable value in the SCM) [12] further showed th… view at source ↗

**Figure 6.** Figure 6: Assembly-stability heatmap after formation rounds. view at source ↗

**Figure 8.** Figure 8: Readout comparison summary. The selected reporting regime is the adaptive soft schedule (adaptive_soft, warm_beta = 0.09, ramp_steps = 20), used consistently in the Alzheimer adaptive experiments and robustness analyses. Explicitly, this is one continuous adaptive gain schedule: directional gain starts at a conservative value (bind_beta = warm_beta) while overlap is monitored each round, and then smoothly … view at source ↗

**Figure 7.** Figure 7: DIRECT mechanism overview. Clarification: The mechanism does not remove reverse links. For every supervised pair (u -> v), both (u -> v) and (v -> u) pathways exist; DIRECT increases forward plasticity temporarily, so reverse remains present but weaker. The directional signal emerges from the asymmetry in strength between forward and reverse pathways, not from the presence or absence of connections. Temp… view at source ↗

**Figure 9.** Figure 9: R3 robustness summary (mean dprop with 95 percent CI across view at source ↗

**Figure 10.** Figure 10: Ground-truth vs produced Alzheimer DAG. Interventional check. Interventional effects were validated using covariate adjustment (conditioning on an admissible set Z to block confounding paths). The estimated average treatment effect (ATE) matched the true interventional ATE in sign and magnitude. Counterfactual check. Counterfactual evaluation (asking what would happen to the same unit under a different tr… view at source ↗

read the original abstract

Can Neural Assemblies -- groups of neurons that fire together and strengthen through co-activation -- learn the direction of causal influence between variables? While established as a computationally general substrate for classification, parsing, and planning, neural assemblies have not yet been shown to internalize causal directionality. We demonstrate that the inherent operations of neural assemblies -- projection, local plasticity control, and sparse winner selection -- are sufficient for directional learning. We introduce DIRECT (DIRectional Edge Coupling/Training), a mechanism that co-activates source and target assemblies under an adaptive gain schedule to internalize directed relations. Unlike backpropagation-based methods, DIRECT relies solely on local plasticity, making the resulting causal claims auditable at the mechanism level. Our findings are verified through a dual-readout validation strategy: (i) synaptic-strength asymmetry, measuring the emergent weight gap between forward and reverse links, and (ii) functional propagation overlap, quantifying the reliability of directional signal flow. Across multiple domains, the framework achieves perfect structural recovery under a supervised, known-structure setting. These results establish neural assemblies as an auditable bridge between biologically plausible dynamics and formal causal models, offering an "explainable by design" framework where causal claims are traceable to specific neural winners and synaptic asymmetries.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper shows neural assemblies can encode known causal directions via local Hebbian updates when an external schedule supplies the directions, but it does not test discovery of those directions from observational data.

read the letter

The core result is that a simple co-activation rule with adaptive gain can produce asymmetric synaptic weights and directional signal flow in neural assemblies. This works under a supervised regime where the experimenter already knows the causal graph and uses it to decide which assemblies to pair as source and target. The authors call the rule DIRECT and validate it with two readouts: the gap in forward versus reverse weights, and how reliably activity propagates in the trained direction. Both readouts come out clean across their test domains, and the whole thing stays local—no backprop required. That is the concrete advance: a biologically motivated substrate that can store directed relations in a way that is traceable to specific winners and weight differences. It is a modest but clear step beyond standard Hebbian learning for undirected associations. The limitation is exactly what the stress-test flags. Because the gain schedule is set from the known structure, the experiment demonstrates faithful encoding rather than inference. The paper never shows the same operations recovering directions when only joint observations are available and no external signal tells the network which variable influences which. The dual validation therefore confirms that imposed directions can be stored asymmetrically, but it leaves the harder causal-discovery case untested. The circularity risk is real but secondary; the main issue is scope. Readers working on neuromorphic hardware or explainable causal models will find the local-plasticity angle useful and the validation approach straightforward to replicate. The work is coherent on its own terms and cites the relevant assembly and causal literature without obvious gaps. It deserves a serious referee who can check the implementation details and ask for an explicit unsupervised or observational-data experiment in revision. I would send it out rather than desk-reject.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces DIRECT, a mechanism enabling neural assemblies to internalize directed causal relations via projection, local plasticity control, and sparse winner selection. Source and target assemblies are co-activated under an adaptive gain schedule derived from known causal structure; the resulting weights are validated by two readouts—emergent synaptic-strength asymmetry and directional propagation overlap—yielding perfect structural recovery in a supervised, known-structure regime. The work positions this as a biologically local, auditable alternative to backpropagation for causal modeling.

Significance. If the central mechanism proves robust beyond the supervised regime, the result would supply a concrete, locally computable bridge between neural-assembly dynamics and formal causal graphs, with built-in auditability at the level of individual synaptic asymmetries and winner selections. The dual-readout validation and emphasis on parameter-light local plasticity constitute genuine strengths that could inform both computational neuroscience and explainable causal AI.

major comments (2)

[Abstract] Abstract: the claim that neural-assembly operations are 'sufficient for directional learning' of causal influence is load-bearing for the title and opening question, yet the reported experiments operate exclusively in a supervised known-structure setting in which the adaptive gain schedule is set using prior knowledge of which assemblies are sources versus targets. This setup demonstrates encoding of externally supplied directions rather than inference of unknown directions from observational data, directly limiting the scope of the central claim.
[Validation strategy] Validation strategy (abstract and methods): both synaptic-strength asymmetry and functional propagation overlap are defined with respect to the same co-activation schedule used to train the weights. This creates a moderate risk that the reported directionality is partly tautological with the training rule rather than an independent test of learned causality; explicit controls or an observational-data ablation are needed to substantiate the dual-readout as evidence of genuine directional internalization.

minor comments (2)

The abstract states 'across multiple domains' without enumeration; the results section should list the domains, any domain-specific parameter choices, and the precise criteria for 'perfect structural recovery' (e.g., exact recovery rate, error bars, exclusion rules).
The adaptive gain schedule is listed as the sole free parameter; its functional form, initialization, and adaptation rule should be given explicitly (ideally with pseudocode or an equation) to support reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the detailed and constructive feedback on our manuscript. We have carefully considered the major comments and provide point-by-point responses below. We believe these revisions will clarify the scope and strengthen the validation of our proposed DIRECT mechanism.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that neural-assembly operations are 'sufficient for directional learning' of causal influence is load-bearing for the title and opening question, yet the reported experiments operate exclusively in a supervised known-structure setting in which the adaptive gain schedule is set using prior knowledge of which assemblies are sources versus targets. This setup demonstrates encoding of externally supplied directions rather than inference of unknown directions from observational data, directly limiting the scope of the central claim.

Authors: We agree that the current experiments demonstrate the ability of neural assemblies to encode directed causal relations when provided with a co-activation schedule derived from known causal structure, rather than inferring directions from purely observational data. The claim of sufficiency refers to the operations (projection, local plasticity, sparse winner selection) being adequate to internalize directionality under this supervised regime, which is a necessary first step before tackling unsupervised inference. This is explicitly stated in the manuscript as 'under a supervised, known-structure setting.' To address this, we will revise the abstract and introduction to more precisely delineate the scope, emphasizing that DIRECT provides a mechanism for directional encoding given causal knowledge, positioning it as a building block for future causal inference in neural systems. We will also add a discussion section on extending to observational settings. revision: yes
Referee: [Validation strategy] Validation strategy (abstract and methods): both synaptic-strength asymmetry and functional propagation overlap are defined with respect to the same co-activation schedule used to train the weights. This creates a moderate risk that the reported directionality is partly tautological with the training rule rather than an independent test of learned causality; explicit controls or an observational-data ablation are needed to substantiate the dual-readout as evidence of genuine directional internalization.

Authors: We appreciate this concern regarding potential circularity in the validation. The synaptic-strength asymmetry is measured post-training as the difference in weights between forward and reverse connections, which emerges from the asymmetric co-activation schedule. The functional propagation overlap tests whether activation in the source assembly reliably propagates to the target under the learned weights, which is a dynamic test separate from the training phase. However, to mitigate the risk of tautology, we will include additional controls in the revised manuscript: (1) a baseline with symmetric co-activation to show that asymmetry does not emerge without directional scheduling, and (2) an ablation using observational data without the adaptive gain schedule to demonstrate that directionality requires the causal-informed training. These will be added to the methods and results sections to provide stronger evidence of genuine internalization. revision: yes

Circularity Check

1 steps flagged

Synaptic asymmetry validation is tautological with the imposed co-activation schedule

specific steps

self definitional [Abstract]
"We introduce DIRECT (DIRectional Edge Coupling/Training), a mechanism that co-activates source and target assemblies under an adaptive gain schedule to internalize directed relations. ... Our findings are verified through a dual-readout validation strategy: (i) synaptic-strength asymmetry, measuring the emergent weight gap between forward and reverse links, and (ii) functional propagation overlap, quantifying the reliability of directional signal flow. Across multiple domains, the framework achieves perfect structural recovery under a supervised, known-structure setting."

The adaptive gain schedule explicitly designates which assemblies are co-activated as source vs. target; local plasticity then strengthens only the forward synapses. The measured 'emergent' synaptic-strength asymmetry is therefore the direct, forced outcome of selective co-activation rather than an independent verification that directionality has been learned; the validation confirms the training rule's effect by construction.

full rationale

The paper's core demonstration is that local Hebbian plasticity under a known-structure co-activation schedule produces asymmetric weights and directional propagation. This is shown to work, but the dual-readout metrics directly measure the consequences of the experimenter-supplied source/target schedule rather than an independent test of direction discovery. The mechanism itself (projection, local plasticity, winner selection) is non-circular and auditable, yet the interpretation of 'directional learning' and 'structural recovery' reduces to successful encoding of externally dictated directions. No self-citation chains or ansatzes are load-bearing; the circularity is limited to the validation step being by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

Only the abstract is available, so the ledger reflects claims made there. The paper introduces DIRECT as a new training schedule and relies on the sufficiency of three standard neural-assembly operations.

free parameters (1)

adaptive gain schedule
Controls the timing and strength of co-activation between source and target assemblies; its exact functional form and hyperparameters are not specified in the abstract.

axioms (1)

domain assumption Projection, local plasticity control, and sparse winner selection are sufficient to internalize directed relations when paired with the DIRECT schedule.
Stated directly in the abstract as the core demonstration.

invented entities (1)

DIRECT mechanism no independent evidence
purpose: Co-activates source and target assemblies under an adaptive gain schedule to produce directional synaptic asymmetry.
Newly introduced training procedure whose details are not provided beyond the name and high-level description.

pith-pipeline@v0.9.0 · 5514 in / 1297 out tokens · 45964 ms · 2026-05-07T10:35:25.322607+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

16 extracted references · 11 canonical work pages

[1]

Guo-Qiang Bi and Mu-Ming Poo. 1998. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type.Journal of Neuroscience18, 24 (1998), 10464–10472. doi:10.1523/JNEUROSCI.18-24-10464.1998

work page doi:10.1523/jneurosci.18-24-10464.1998 1998
[2]

György Buzsáki. 2010. Neural syntax: cell assemblies, synapsembles, and readers. Neuron68, 3 (2010), 362–385. doi:10.1016/j.neuron.2010.10.023

work page doi:10.1016/j.neuron.2010.10.023 2010
[3]

Max Dabagia, Daniel Mitropolsky, Christos Papadimitriou, and Santosh Vempala
[4]

arXiv:2406.07715 [q-bio.NC] doi:10.48550/arXiv.2406.07715

Coin-flipping in the brain: statistical learning with neuronal assemblies. arXiv:2406.07715 [q-bio.NC] doi:10.48550/arXiv.2406.07715

work page doi:10.48550/arxiv.2406.07715
[5]

Max Dabagia, Christos Papadimitriou, and Santosh Vempala. 2022. Neural As- semblies of neurons learn to classify well-separated distributions. InProceedings of The 35th Conference on Learning Theory (Proceedings of Machine Learning Research, Vol. 178). PMLR, Online, 4705–4744. https://proceedings.mlr.press/v1 78/dabagia22a.html

2022
[6]

Max Dabagia, Christos Papadimitriou, and Santosh Vempala. 2025. Computation with sequences of assemblies in a model of the brain.Neural Computation37, 1 (2025), 193–233. doi:10.1162/neco_a_01720

work page doi:10.1162/neco_a_01720 2025
[7]

Francesco d’Amore, Daniel Mitropolsky, Pierluigi Crescenzi, Emanuele Natale, and Christos Papadimitriou. 2022. Planning with biological neurons and synapses. https://aaai.org/papers/00021- planning- with- biological- neurons- and- synapses/ In Proceedings of the AAAI Conference on Artificial Intelligence

2022
[8]

Donald O. Hebb. 1949.The Organization of Behavior: A Neuropsychological Theory. Wiley, New York, NY, USA. https://openlibrary.org/isbn/9780805843002

work page arXiv 1949
[9]

Jakub Kuzilek, Martin Hlosta, and Zdenek Zdrahal. 2017. Open University Learn- ing Analytics dataset.Scientific Data4 (2017), 170171. doi:10.1038/sdata.2017.171

work page doi:10.1038/sdata.2017.171 2017
[10]

Henry Markram, Joachim Lübke, Michael Frotscher, and Bert Sakmann. 1997. Regulation of synaptic efficacy by coincidence of postsynaptic APs and EPSPs. Science275, 5297 (1997), 213–215. doi:10.1126/science.275.5297.213

work page doi:10.1126/science.275.5297.213 1997
[11]

and Papadimitriou, Christos H

Daniel Mitropolsky, Michael Collins, and Christos H. Papadimitriou. 2021. A biologically plausible parser.Transactions of the Association for Computational Linguistics9 (2021), 1374–1388. doi:10.1162/tacl_a_00432

work page doi:10.1162/tacl_a_00432 2021
[12]

Brain com- putation by assemblies of neurons,

Christos H. Papadimitriou, Santosh Vempala, Daniel Mitropolsky, Michael Collins, and Wolfgang Maass. 2020. Brain computation by neural assemblies of neurons. Proceedings of the National Academy of Sciences117, 25 (2020), 14464–14472. doi:10.1073/pnas.2001893117

work page doi:10.1073/pnas.2001893117 2020
[13]

Cambridge University Press, 2nd edi- tion, 2009

Judea Pearl. 2009.Causality: Models, Reasoning, and Inference(2nd ed.). Cambridge University Press, Cambridge, UK. doi:10.1017/CBO9780511803161

work page doi:10.1017/cbo9780511803161 2009
[14]

2017.Elements of Causal Inference: Foundations and Learning Algorithms

Jonas Peters, Dominik Janzing, and Bernhard Schölkopf. 2017.Elements of Causal Inference: Foundations and Learning Algorithms. MIT Press, Cambridge, MA, USA. https://direct.mit.edu/books/book/4376/Elements-of-Causal-Inference

2017
[15]

Miller, and L

Sen Song, Kenneth D. Miller, and L. F. Abbott. 2000. Competitive Hebbian learning through spike-timing-dependent synaptic plasticity.Nature Neuroscience3, 9 (2000), 919–926. doi:10.1038/78829

work page doi:10.1038/78829 2000
[16]

2000.Causation, Prediction, and Search(2nd ed.)

Peter Spirtes, Clark Glymour, and Richard Scheines. 2000.Causation, Prediction, and Search(2nd ed.). MIT Press, Cambridge, MA, USA. https://direct.mit.edu/b ooks/book/2605/Causation-Prediction-and-Search

2000

[1] [1]

Guo-Qiang Bi and Mu-Ming Poo. 1998. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type.Journal of Neuroscience18, 24 (1998), 10464–10472. doi:10.1523/JNEUROSCI.18-24-10464.1998

work page doi:10.1523/jneurosci.18-24-10464.1998 1998

[2] [2]

György Buzsáki. 2010. Neural syntax: cell assemblies, synapsembles, and readers. Neuron68, 3 (2010), 362–385. doi:10.1016/j.neuron.2010.10.023

work page doi:10.1016/j.neuron.2010.10.023 2010

[3] [3]

Max Dabagia, Daniel Mitropolsky, Christos Papadimitriou, and Santosh Vempala

[4] [4]

arXiv:2406.07715 [q-bio.NC] doi:10.48550/arXiv.2406.07715

Coin-flipping in the brain: statistical learning with neuronal assemblies. arXiv:2406.07715 [q-bio.NC] doi:10.48550/arXiv.2406.07715

work page doi:10.48550/arxiv.2406.07715

[5] [5]

Max Dabagia, Christos Papadimitriou, and Santosh Vempala. 2022. Neural As- semblies of neurons learn to classify well-separated distributions. InProceedings of The 35th Conference on Learning Theory (Proceedings of Machine Learning Research, Vol. 178). PMLR, Online, 4705–4744. https://proceedings.mlr.press/v1 78/dabagia22a.html

2022

[6] [6]

Max Dabagia, Christos Papadimitriou, and Santosh Vempala. 2025. Computation with sequences of assemblies in a model of the brain.Neural Computation37, 1 (2025), 193–233. doi:10.1162/neco_a_01720

work page doi:10.1162/neco_a_01720 2025

[7] [7]

Francesco d’Amore, Daniel Mitropolsky, Pierluigi Crescenzi, Emanuele Natale, and Christos Papadimitriou. 2022. Planning with biological neurons and synapses. https://aaai.org/papers/00021- planning- with- biological- neurons- and- synapses/ In Proceedings of the AAAI Conference on Artificial Intelligence

2022

[8] [8]

Donald O. Hebb. 1949.The Organization of Behavior: A Neuropsychological Theory. Wiley, New York, NY, USA. https://openlibrary.org/isbn/9780805843002

work page arXiv 1949

[9] [9]

Jakub Kuzilek, Martin Hlosta, and Zdenek Zdrahal. 2017. Open University Learn- ing Analytics dataset.Scientific Data4 (2017), 170171. doi:10.1038/sdata.2017.171

work page doi:10.1038/sdata.2017.171 2017

[10] [10]

Henry Markram, Joachim Lübke, Michael Frotscher, and Bert Sakmann. 1997. Regulation of synaptic efficacy by coincidence of postsynaptic APs and EPSPs. Science275, 5297 (1997), 213–215. doi:10.1126/science.275.5297.213

work page doi:10.1126/science.275.5297.213 1997

[11] [11]

and Papadimitriou, Christos H

Daniel Mitropolsky, Michael Collins, and Christos H. Papadimitriou. 2021. A biologically plausible parser.Transactions of the Association for Computational Linguistics9 (2021), 1374–1388. doi:10.1162/tacl_a_00432

work page doi:10.1162/tacl_a_00432 2021

[12] [12]

Brain com- putation by assemblies of neurons,

Christos H. Papadimitriou, Santosh Vempala, Daniel Mitropolsky, Michael Collins, and Wolfgang Maass. 2020. Brain computation by neural assemblies of neurons. Proceedings of the National Academy of Sciences117, 25 (2020), 14464–14472. doi:10.1073/pnas.2001893117

work page doi:10.1073/pnas.2001893117 2020

[13] [13]

Cambridge University Press, 2nd edi- tion, 2009

Judea Pearl. 2009.Causality: Models, Reasoning, and Inference(2nd ed.). Cambridge University Press, Cambridge, UK. doi:10.1017/CBO9780511803161

work page doi:10.1017/cbo9780511803161 2009

[14] [14]

2017.Elements of Causal Inference: Foundations and Learning Algorithms

Jonas Peters, Dominik Janzing, and Bernhard Schölkopf. 2017.Elements of Causal Inference: Foundations and Learning Algorithms. MIT Press, Cambridge, MA, USA. https://direct.mit.edu/books/book/4376/Elements-of-Causal-Inference

2017

[15] [15]

Miller, and L

Sen Song, Kenneth D. Miller, and L. F. Abbott. 2000. Competitive Hebbian learning through spike-timing-dependent synaptic plasticity.Nature Neuroscience3, 9 (2000), 919–926. doi:10.1038/78829

work page doi:10.1038/78829 2000

[16] [16]

2000.Causation, Prediction, and Search(2nd ed.)

Peter Spirtes, Clark Glymour, and Richard Scheines. 2000.Causation, Prediction, and Search(2nd ed.). MIT Press, Cambridge, MA, USA. https://direct.mit.edu/b ooks/book/2605/Causation-Prediction-and-Search

2000