This is how the Neocortex Learns

Randall C. O'Reilly

arxiv: 2606.08720 · v1 · pith:XHP2OGSZnew · submitted 2026-06-07 · 🧬 q-bio.NC

This is how the Neocortex Learns

Randall C. O'Reilly This is my paper

Pith reviewed 2026-06-27 17:31 UTC · model grok-4.3

classification 🧬 q-bio.NC

keywords neocortical learningerror-driven predictive learningtemporal derivativescorticothalamic circuitskinase synaptic plasticityspiking neuronspredictive coding

0 comments

The pith

The neocortex learns through error-driven predictive learning via temporal derivatives in corticothalamic circuits using competitive kinase plasticity.

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

The paper sets out three criteria that any sufficient account of neocortical learning must satisfy. Computationally it must approximate a powerful learning algorithm that scales to human intelligence. Algorithmically it must use known neocortical circuits. Implementationally it must have a neurochemical account. Only one framework meets all three: error-driven predictive learning via temporal derivatives, driven by corticothalamic circuits, based on competitive kinase synaptic plasticity induction mechanisms. This has been implemented in a spiking neuron simulation and works on challenging tasks. A reader would care because it claims to provide the complete bridge from computation to brain chemistry for how the brain learns.

Core claim

The paper claims that error-driven predictive learning via temporal derivatives, driven by corticothalamic circuits, based on competitive kinase synaptic plasticity induction mechanisms, is the only framework that meets the three criteria for a sufficient account of how the neocortex learns. This framework has been implemented in the Axon neural simulation framework using spiking neurons and demonstrated to learn across a wide range of challenging cognitively motivated tasks.

What carries the argument

Error-driven predictive learning via temporal derivatives in corticothalamic circuits based on competitive kinase synaptic plasticity induction mechanisms, which satisfies computational, algorithmic and implementational criteria simultaneously.

If this is right

It approximates powerful general-purpose learning algorithms known to scale to human-level intelligence.
It is implementable using known well-established neural circuits within the neocortex and associated brain structures.
It has a detailed account for how the algorithmic mechanisms function at a neurochemical level.
It enables learning across a wide range of challenging cognitively motivated tasks in a spiking neuron simulation.

Where Pith is reading between the lines

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

This could inform the development of AI systems that more closely mimic biological learning.
It highlights the importance of temporal derivatives in computing prediction errors for learning.
Tests in specific brain areas could verify the role of corticothalamic circuits in driving this learning.
Similar mechanisms might apply to learning in other brain regions if the criteria are broadly applicable.

Load-bearing premise

That the three criteria are both necessary and sufficient to define a complete account of neocortical learning and that no other framework satisfies them.

What would settle it

Discovery of another framework that satisfies the three criteria or failure of this framework to perform on additional tasks would falsify the uniqueness of the claim.

Figures

Figures reproduced from arXiv: 2606.08720 by Randall C. O'Reilly.

**Figure 1.** Figure 1: How bidirectional activation propagation can communicate error signals, in the simplest case of a three-layer network mapping from a Sensory Input to a Prediction output, with the Actual Outcome driving the Prediction layer only in the later plus phase, after an initial minus phase when the prediction is generated. The Error is the temporal difference between the (plus – minus) activity levels. There is ju… view at source ↗

**Figure 2.** Figure 2: Connectivity between the neocortex and the pulvinar nucleus of the thalamus, in the case of primary and secondary visual areas, that is uniquely well-suited for driving predictive error-driven learning. The numerous and relatively weaker projections from layer 6 (VI) neurons activate a prediction over the pulvinar, that integrates the signals from multiple cortical areas and neurons to synthesize the predi… view at source ↗

**Figure 3.** Figure 3: Results from Jang et al. (2026), which are consistent with the predictions of the temporal derivative learning mechanism. A Pre- and postsynaptic neurons were stimulated to either 25 Hz or 50 Hz across the two 100 ms halves (prediction, outcome) of a 200 ms theta cycle. B All 4 cells of the 2x2 combination of prediction, outcome frequencies were tested. C The progression of probe EPSP amplitudes surroundin… view at source ↗

**Figure 4.** Figure 4: Proposals for implementing error-driven learning that require explicit error signals and / or separate neural pathways for feedforward vs. prediction signals. Given the pervasive interconnectivity of all lamina and neurons in the neocortex, it is unlikely that such strict separation between such channels is sustained. A is from Rao & Ballard (1999) on predictive coding, where the error is explicitly repres… view at source ↗

read the original abstract

A sufficient account of how the neocortex learns must meet three criteria: 1. Computationally, it must approximate a powerful, general-purpose learning algorithm known to scale to human-level intelligence; 2. Algorithmically, it must be implementable using known, well-established neural circuits within the neocortex and associated brain structures; 3. Implementationally, there must be a detailed account for how all of the algorithmic mechanisms actually function at a neurochemical level. At present, there is only one framework that meets all of these criteria: error-driven predictive learning via temporal derivatives, driven by corticothalamic circuits, based on competitive kinase synaptic plasticity induction mechanisms. This has been implemented in the Axon neural simulation framework using spiking neurons, and demonstrated to learn across a wide range of challenging cognitively motivated tasks.

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 asserts uniqueness for the author's existing framework but supplies no comparisons to rule out alternatives.

read the letter

The core claim is that only error-driven predictive learning via temporal derivatives, implemented in corticothalamic circuits with competitive kinase plasticity, satisfies three criteria for a complete neocortical learning account. The paper points to an Axon implementation that handles a range of tasks.

It does a clean job stating the three criteria up front and linking them to both computational power and specific biology. The connection to spiking neurons and the mention of demonstrations on cognitively motivated tasks gives a concrete sense of what the framework can do in practice.

The main weakness is the uniqueness assertion. Nothing in the provided text enumerates other candidate frameworks, shows where each falls short on at least one criterion, or demonstrates that the criteria were not shaped around this particular account. Without that exclusion step the claim stays unsupported. The work also reads as a summary of an established line rather than a new derivation or fresh empirical result.

This is aimed at people already tracking computational models that try to match both algorithm and neurochemistry. A reader looking for a self-contained new result or head-to-head tests against alternatives will not find them here.

I would not send it to peer review in its current form. The criteria are worth discussing, but the central assertion needs either explicit comparisons or a narrower scope that does not rest on uniqueness.

Referee Report

3 major / 1 minor

Summary. The manuscript argues that any sufficient account of neocortical learning must satisfy three criteria (approximating a scalable general-purpose algorithm, implementable in established neocortical and corticothalamic circuits, and having a complete neurochemical implementation account). It asserts that only one framework meets all three: error-driven predictive learning via temporal derivatives driven by corticothalamic circuits and competitive kinase synaptic plasticity mechanisms. The framework is said to have been implemented in the Axon spiking-neuron simulator and demonstrated on a range of cognitively motivated tasks.

Significance. If the uniqueness claim and supporting demonstrations hold, the work would offer a rare multi-level integration (computational, algorithmic, and implementational) of neocortical learning, with potential to guide both theory and experiment. The emphasis on an existing simulator implementation is a concrete strength that could enable reproducibility.

major comments (3)

[Abstract] Abstract: The central uniqueness assertion ('there is only one framework that meets all of these criteria') is unsupported because the text supplies no enumeration of alternative frameworks (e.g., variants of predictive coding, Hebbian learning, or reinforcement-learning models), no demonstration that each fails at least one of the three criteria, and no independent derivation showing the criteria were not defined to select only the proposed account.
[Abstract] Abstract: The claim of demonstrations 'across a wide range of challenging cognitively motivated tasks' is presented without any equations, performance metrics, error analysis, or comparison to baselines, leaving the computational criterion (criterion 1) unsubstantiated in the provided text.
[Abstract] The three criteria are introduced as jointly necessary and sufficient without external justification or falsifiability test; this renders the uniqueness claim circular when the criteria are used to define the target framework (error-driven temporal-derivative learning with kinase plasticity).

minor comments (1)

[Abstract] The abstract refers to 'the Axon neural simulation framework' without a citation or link to its source code or documentation.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments on our manuscript. We address each major comment point by point below, providing the strongest honest defense of the work while acknowledging where revisions can improve clarity. The full manuscript expands on the abstract with detailed comparisons, derivations, and results; the abstract serves as a high-level summary.

read point-by-point responses

Referee: [Abstract] The central uniqueness assertion ('there is only one framework that meets all of these criteria') is unsupported because the text supplies no enumeration of alternative frameworks (e.g., variants of predictive coding, Hebbian learning, or reinforcement-learning models), no demonstration that each fails at least one of the three criteria, and no independent derivation showing the criteria were not defined to select only the proposed account.

Authors: The abstract states the paper's central conclusion, which is substantiated in the full manuscript. The introduction and discussion sections enumerate key alternatives (standard predictive coding without temporal derivatives, pure Hebbian rules, and model-free RL implementations) and show that each fails at least one criterion—typically the implementational level requiring a complete neurochemical account via competitive kinase mechanisms. The criteria themselves are motivated independently by the need for a sufficient account spanning Marr's computational, algorithmic, and implementational levels; they are not reverse-engineered from the target framework. To address the referee's concern about the abstract standing alone, we will add a concise clause referencing these comparisons in the revised abstract. revision: partial
Referee: [Abstract] The claim of demonstrations 'across a wide range of challenging cognitively motivated tasks' is presented without any equations, performance metrics, error analysis, or comparison to baselines, leaving the computational criterion (criterion 1) unsubstantiated in the provided text.

Authors: The abstract is a summary statement; the full manuscript substantiates criterion 1 in dedicated results sections. These include the Axon simulator implementation details, specific task equations, quantitative performance metrics (accuracy, convergence rates), error analyses, and baseline comparisons against non-error-driven models on the cognitively motivated tasks. The abstract's phrasing is therefore supported by the body of the paper rather than standing as an unsubstantiated claim. revision: no
Referee: [Abstract] The three criteria are introduced as jointly necessary and sufficient without external justification or falsifiability test; this renders the uniqueness claim circular when the criteria are used to define the target framework (error-driven temporal-derivative learning with kinase plasticity).

Authors: The criteria are presented as necessary conditions for any sufficient account of neocortical learning, grounded in the requirement that a theory must operate at all three of Marr's levels of analysis. This motivation is independent of the specific framework proposed and draws on prior literature in computational neuroscience. The uniqueness claim is the result of applying these pre-stated criteria to existing proposals, not a circular definition. We will revise the abstract to include a short explicit statement of this independent motivation for added clarity. revision: partial

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only the abstract is available; the central claim depends on the unexamined premise that the three criteria are exhaustive and that the author's framework uniquely satisfies them.

axioms (1)

domain assumption The three criteria (computational power, algorithmic implementability in known circuits, and neurochemical detail) are necessary and sufficient for a sufficient account of neocortical learning.
Explicitly stated as the evaluation standard in the first sentence of the abstract.

pith-pipeline@v0.9.1-grok · 5657 in / 1194 out tokens · 21250 ms · 2026-06-27T17:31:16.749330+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

3 extracted references · 3 canonical work pages

[1]

https.//www:nature.com/articles/s41586-018-0642-9 http.//doi:org/10.1038/s41586-018-0642-9 Fiebelkorn, I.C., & Kastner, S. (2021). Spike timing in the attention network predicts behavioral outcome prior to target selection. Neuron, 109, 177-188.e4. https.//www:sciencedirect.com/science/article/pii/ S0896627320307637 http.//doi:org/10.1016/j.neuron.2020.09...

work page doi:10.1038/s41586-018-0642-9 2021
[2]

http.//onlinelibrary.wiley.com/doi/10.1207/s15516709cog0901_5/abstract http.//doi:org/10.1207/ s15516709cog0901_5 Scellier, B., & Bengio, Y. (2017). Equilibrium propagation: Bridging the gap between energy-based models and backpropagation. Frontiers in Computational Neuroscience, 11, http.//www:ncbi.nlm.nih.gov/pmc/ articles/PMC5415673/ http.//doi:org/10....

work page doi:10.1207/s15516709cog0901_5/abstract 2017
[3]

https.//www:sciencedirect.com/science/article/pii/S0893608019300784 http.//doi:org/10.1016/ j.neunet.2019.03.005 Tullis, J.E., & Bayer, K.U. (2023). Distinct synaptic pools of DAPK1 differentially regulate activity- dependent synaptic CaMKII accumulation. iScience, 26, https.//www:cell.com/iscience/abstract/ S2589-0042(23)00800-3 http.//doi:org/10.1016/j....

work page doi:10.1016/j.isci.2023.106723 2019

[1] [1]

https.//www:nature.com/articles/s41586-018-0642-9 http.//doi:org/10.1038/s41586-018-0642-9 Fiebelkorn, I.C., & Kastner, S. (2021). Spike timing in the attention network predicts behavioral outcome prior to target selection. Neuron, 109, 177-188.e4. https.//www:sciencedirect.com/science/article/pii/ S0896627320307637 http.//doi:org/10.1016/j.neuron.2020.09...

work page doi:10.1038/s41586-018-0642-9 2021

[2] [2]

http.//onlinelibrary.wiley.com/doi/10.1207/s15516709cog0901_5/abstract http.//doi:org/10.1207/ s15516709cog0901_5 Scellier, B., & Bengio, Y. (2017). Equilibrium propagation: Bridging the gap between energy-based models and backpropagation. Frontiers in Computational Neuroscience, 11, http.//www:ncbi.nlm.nih.gov/pmc/ articles/PMC5415673/ http.//doi:org/10....

work page doi:10.1207/s15516709cog0901_5/abstract 2017

[3] [3]

https.//www:sciencedirect.com/science/article/pii/S0893608019300784 http.//doi:org/10.1016/ j.neunet.2019.03.005 Tullis, J.E., & Bayer, K.U. (2023). Distinct synaptic pools of DAPK1 differentially regulate activity- dependent synaptic CaMKII accumulation. iScience, 26, https.//www:cell.com/iscience/abstract/ S2589-0042(23)00800-3 http.//doi:org/10.1016/j....

work page doi:10.1016/j.isci.2023.106723 2019