Learning complex temporal dependencies via local synaptic plasticity

Temporal predictive coding (tPC) shows structurally equivalent local Hebbian weight updates to tBPTT1, enabling learning of temporal dependencies in recurrent networks without backpropagation through time.

• Why it matters: Backpropagation through time is widely used to model neural and behavioral learning, but its biological plausibility is limited, leaving efficient temporal learning principles unresolved.
• What they did: The work analyzes and extends tPC, relating it to BPTT, reservoir computing, and e-prop, and demonstrates equivalence to tBPTT1 with one-step gradient propagation into the past.
• The result: tPC leverages reservoir dynamics for short-range temporal context, uses recurrent state-space trajectory shaping for readout, and—when augmented with eligibility traces—solves temporally extended context-dependent tasks.

AI summary of the abstract · may contain errors · not medical advice · read the paper.