#32 · Systems biology

Simple and recurrent neural models decode vision and mimic primate decision dynamics

The dominant thread is how far simple computational models recover perceptual content and behavioral structure from neural activity. In primate intracortical spiking data, a temporal-attention plus MLP decoder reaches 70% top-1 retrieval of visual images, showing that substantial visual information is extractable with a lightweight readout. Separately, recurrent networks trained with deep reinforcement learning on a noisy task reproduce primate-like speed-accuracy tradeoffs and changes of mind, with neural-like dynamics. A published study of mouse visual cortex examines higher-level spatial prediction within predictive processing theory, concerning how spatial and temporal context shape sensory prediction. The new element across these items is the joint treatment of decoding and decision-level computation within trained models, though the evidence base here is limited to three items.