🔗 Source: arXiv

This is how the Neocortex Learns

🚀 Technical Novelty

  • Mechanism: Error-driven predictive learning via temporal derivatives (phase-difference activation) combined with corticothalamic circuits and competitive kinase synaptic plasticity, implemented in the Axon spiking neural framework.
  • Nuance: Replaces explicit error pathways and segregated feedforward/feedback channels with bidirectional excitatory networks that implicitly encode gradients as temporal activity differences, aligning with known neocortical microcircuitry rather than artificial backpropagation architectures.

💡 Yield

  • Demonstrates that temporal derivative algorithms mathematically approximate error backpropagation while satisfying computational, algorithmic, and implementational criteria for biological plausibility.
  • Integrates behavioral timescale synaptic plasticity (BTSP) and eligibility traces to resolve the temporal credit assignment problem in deep layered networks without requiring instantaneous gradient signals.

⚠️ Limitations

  • The precise driving target signal for layer 5 neocortical plasticity remains empirically unresolved.
  • Relies on theoretical mapping between abstract algorithms and complex neurochemical processes, requiring further in vivo validation.
  • Focuses on computational neuroscience theory rather than large-scale AI engineering benchmarks or architectural innovations.