Joint encoding of "what" and "when" predictions through error-modulated plasticity in biologically-plausible spiking networks

Neuroscience researchers Yohei Yamada and Zenas C. Chao have published a breakthrough paper demonstrating how a single population of spiking neurons can learn to make complete predictions about future events. Their model, detailed in arXiv:2510.14382, addresses a key limitation in existing computational neuroscience: most models either predict what will happen OR when it will happen, but not both simultaneously with probability estimates. The researchers show that a heterogeneous Izhikevich spiking reservoir can acquire and flexibly maintain what they term a 'complete prediction object'—jointly specifying identity, timing, and likelihood of future events.

The technical innovation centers on an error-modulated, attention-gated three-factor Hebbian learning rule that operates locally without requiring biologically implausible global error broadcasts. When tested on tasks that independently manipulate event identity, latency, and probability, the network developed time-locked anticipatory activity whose amplitude scaled with outcome probability. Crucially, identity and timing components self-organized into near-orthogonal readout subspaces within the same neural population, demonstrating multidimensional predictive structure can emerge without anatomical modularization. Compared to traditional least-squares approaches, this local gated plasticity enables stable recalibration under nonstationary conditions, suggesting cortical mixed-selective populations with neuromodulator-gated plasticity may be sufficient for flexible predictive cognition.