#26 · Systems biology

Recurrent network models link memory efficiency and cortical dynamics to perception

The four preprints share a recurrent-network framing of temporal neural computation, though they address different levels of description. One study shows that memory solutions in RNNs are partitioned by efficiency and robustness constraints, with noise robustness opposing activity efficiency, and models premotor cortex. A second uses a brain-inspired recurrent network to model how cortical dynamics group objects in natural scenes and predicts human reaction times. A third proposes a dual-stream recurrent graph model in which grid-like, context-invariant codes emerge along the dorsal visual stream when trained on Something-Something V2. The fourth recasts neural decoding as Bayesian cognitive inference with brain-intrinsic priors. What is new across the set is the use of explicit computational constraints and behavioral or neural benchmarks to account for how representations evolve over time and generalize.