O'Reilly preprint: mammalian cortex approximates backpropagation via 200-millisecond theta cycles

Randall C. O'Reilly's preprint This is How the Neocortex Learns argues that the mammalian neocortex approximates backpropagation through a "temporal derivative model" — error gradients encoded as the difference between successive activation states across a 200-millisecond theta cycle. Released on arXiv this week (arxiv.org/abs/2606.08720), the work synthesizes decades of neuroscience, electrophysiology, and computational theory to resolve a longstanding debate: whether biological brains can perform credit assignment at the scale of deep networks.

The core claim rests on bidirectional corticothalamic loops that drive prediction-outcome comparisons, with subcellular kinase-mediated synaptic plasticity implementing gradient updates — all without dedicated "error neurons" or biologically implausible backward passes. O'Reilly frames this as a unified learning theory for mammalian cortex, bridging artificial gradient descent and the sparse, event-driven signaling of real neurons.

Neuromorphic hardware implications

The paper offers a concrete blueprint for on-chip learning rules in energy-constrained neuromorphic systems. By showing how temporal differences can substitute for explicit error signals, the model points toward hardware architectures that scale like deep nets but run on the sparse, asynchronous dynamics of biological circuits. Competitive synaptic plasticity — kinases competing for limited resources at individual synapses — naturally regularizes learning without external weight decay, O'Reilly notes.

No code or trained model accompanies the preprint; the work is purely theoretical. A detailed review is available on ArxivIQ Substack. The paper lands amid renewed interest in biologically plausible alternatives to standard backprop, though its 200-millisecond timescale and reliance on thalamic gating remain open questions for experimental validation.