A collaboration between SISSA’s Physics and Neuroscience groups has shed new light on how memories are stored and retrieved in the brain, unifying decades of behavioral and theoretical research. The study, led by professors Sebastian Goldt and Mathew E. Diamond and first-authored by Francesca Schönsberg (now at the École Normale Supérieure), has just been published in Neuron.
Perceptual memory – our ability to extract, store, and use information from the sensory world – has long intrigued scientists. Yet the field remains fragmented: researchers often build separate models around specific tasks. The SISSA team set out to bridge these gaps by developing a single framework capable of explaining diverse forms of perceptual memory.
Their key insight came from studying perceptual biases – systematic distortions that reveal how the brain’s internal representations deviate from the external world’s physical reality. When a neural network displays the same distortions seen in behavior, it likely captures a core mechanism of perceptual memory itself.
Perception of current stimuli shows two opposite tendencies: contraction toward past experiences and repulsion away from them. By combining computational modeling with data from humans and rodents, the SISSA team showed that both effects can emerge naturally from a single mechanism. A recurrent neural network governed by Hebbian plasticity (“cells that fire together wire together”) reproduced experimental data in three distinct paradigms – working memory, reference memory, and a novel “one-back” task, without task-dependent fine-tuning. What differed was not the memory mechanism but the way the network’s activity was read out, revealing how a single memory network can flexibly support different cognitive demands.