| name | neocortex-error-driven-predictive-learning |
| description | Neocortex learning framework via error-driven predictive learning with temporal derivatives, corticothalamic circuits, and competitive kinase synaptic plasticity. Three-criteria account of neocortex learning: computational, algorithmic, and implementational. Activation: neocortex learning, predictive coding, error-driven learning, corticothalamic circuits, synaptic plasticity, temporal derivatives. |
Context
Paper: arXiv:2606.08720 - "This is how the Neocortex Learns"
Authors: Randall C. O'Reilly
Submitted: 7 Jun 2026
Category: Neurons and Cognition (q-bio.NC)
Problem
A sufficient account of neocortex learning must meet three criteria:
- Computationally: Must approximate a powerful, general-purpose learning algorithm that scales to human-level intelligence
- Algorithmically: Must be implementable using known, well-established neural circuits within neocortex
- Implementationally: Must have detailed neurochemical mechanisms at molecular level
Core Methodology
Error-driven predictive learning via temporal derivatives meets all three criteria:
1. Computational Level
Temporal difference learning approximates backpropagation:
- Error signal = derivative of activation over time: $\delta = \frac{dA}{dt}$
- Prediction = current activation $A(t)$
- Target = future activation $A(t+\Delta t)$
- Error = $A(t+\Delta t) - A(t)$ (temporal derivative approximation)
This implements a form of predictive coding where:
- Cortex generates predictions about future inputs
- Errors drive learning when predictions fail
- Temporal derivatives provide error signals without explicit backpropagation
2. Algorithmic Level
Corticothalamic circuits implement the algorithm:
Thalamus (prediction generator)
↓ sends predictions
Cortex (error detector)
↓ computes temporal derivative
↓ sends error signals
Thalamus (error integrator)
↓ updates predictions
↓ drives synaptic plasticity
Key circuit mechanisms:
- Layer 6 corticothalamic projections: Generate predictions
- Layer 4 thalamocortical inputs: Provide actual inputs
- Temporal comparison: Layer 4 computes prediction error
- Feedback pathway: Error signals propagate back through Layer 5/6
3. Implementational Level
Competitive kinase synaptic plasticity mechanisms:
-
CaMKII vs PKC competition:
- CaMKII activated by NMDA receptor calcium influx (LTP pathway)
- PKC activated by error signals ( LTD pathway)
- Competition determines synaptic weight change direction
-
Temporal derivative encoding:
- Early calcium influx (prediction phase) → CaMKII dominance → LTP
- Late calcium influx (error phase) → PKC dominance → LTD
- Net weight change = LTP - LTD (temporal derivative)
-
Neurochemical cascade:
Prediction phase: NMDA → Ca²⁺ → CaMKII → GluR1 phosphorylation → AMPA insertion (LTP)
Error phase: Error signal → PKC → GluR2 phosphorylation → AMPA removal (LTD)
Net: Weight change = Δ(AMPA insertion) - Δ(AMPA removal)
Implementation in Axon Framework
Spiking neural network simulation demonstrates learning:
class NeocortexLayer:
def __init__(self):
self.prediction_neurons = Layer6Neurons()
self.error_neurons = Layer4Neurons()
self.thalamic_input = ThalamicProjection()
def learn(self, input_spike_train, target_spike_train):
prediction = self.prediction_neurons.predict(input_spike_train)
actual = self.thalamic_input.receive(target_spike_train)
error = temporal_derivative(actual, prediction)
for synapse in self.synapses:
if error > threshold:
synapse.ltd(pkc_activation)
else:
synapse.ltp(camkii_activation)
Key Results
- Demonstrated learning on challenging cognitively motivated tasks
- Three-criteria framework satisfies computational, algorithmic, and implementational levels
- Error-driven predictive learning = backpropagation approximation via temporal derivatives
- Corticothalamic circuits provide anatomical substrate
- Competitive kinase plasticity provides molecular mechanism
Implementation Steps
- Initialize Axon spiking network (Layer 4, 5, 6 + thalamic projections)
- Define prediction pathway: Layer 6 → thalamus → Layer 4
- Define error computation: Temporal derivative in Layer 4
- Configure synaptic plasticity: CaMKII vs PKC competition
- Train on task: Input → prediction → error → weight update
- Verify learning: Compare to backpropagation performance
Pitfalls
- No thalamic circuit: Cannot compute temporal derivatives without thalamus
- Wrong temporal window: Derivative requires precise timing (prediction vs error phase)
- Missing kinase competition: Only LTP or only LTD fails → need both
- Non-spiking implementation: Continuous activation loses temporal structure
- No predictive coding framework: Error-driven learning requires prediction generation
Verification
prediction = layer6.generate_prediction(input)
actual = thalamus.receive(input)
error = actual - prediction
for synapse in error_neurons.synapses:
weight_change = synapse.weight - initial_weight
if error > 0:
assert weight_change < 0
else:
assert weight_change > 0
bp_error = backpropagate(target, output)
td_error = temporal_derivative(output, prediction)
assert correlation(bp_error, td_error) > 0.8
Applications
- Brain-inspired AI (error-driven learning without backpropagation)
- Neural circuit modeling (corticothalamic loops)
- Synaptic plasticity simulation (kinase competition)
- Cognitive task learning (predictive coding framework)
References
- Paper: https://arxiv.org/abs/2606.08720
- Axon framework: O'Reilly et al. spiking neural simulation
- Related: Predictive coding, Temporal difference learning, CaMKII/PKC plasticity
Activation Keywords
neocortex learning, predictive coding, error-driven learning, corticothalamic circuits, synaptic plasticity, temporal derivatives, competitive kinase, CaMKII PKC, backpropagation approximation, Axon framework