Ravi Srinivasan

"Forward-only" algorithms, which train neural networks while avoiding a backward pass, have recently gained attention as a way of solving the biologically unrealistic aspects of backpropagation. Here, we first address compelling challenges related to the "forward-only" rules, which include reducing the performance gap with backpropagation and providing an analytical understanding of their dynamics. To this end, we show that the forward-only algorithm with top-down feedback is well-approximated by an "adaptive-feedback-alignment" algorithm, and we analytically track its performance during learning in a prototype high-dimensional setting. Then, we compare different versions of forward-only algorithms, focusing on the Forward-Forward and PEPITA frameworks, and we show that they share the same learning principles. Overall, our work unveils the connections between three key neuro-inspired learning rules, providing a link between "forward-only" algorithms, i.e., Forward-Forward and PEPITA, and an approximation of backpropagation, i.e., Feedback Alignment.

The ability to transiently remember what happened where and when is a cornerstone of cognitive function. Forming and recalling working memories depends on detecting novelty, building associations to prior knowledge, and dynamically retrieving context-relevant information. Previous studies have scrutinized the neural machinery for individual components of recognition or associative memory under laboratory conditions, such as recalling elements from arbitrary lists of words or pictures. In this study, we implemented a well-known card- matching game that integrates multiple components of memory formation together in a naturalistic setting to investigate the dynamic neural processes underlying complex natural human memory. We recorded intracranial field potentials from 1,750 depth or subdural electrodes implanted in 20 patients with pharmacologically-intractable epilepsy while they were performing the task. We leveraged generalized linear models to simultaneously assess the relative contribution of neural responses to distinct task components. Neural activity in the gamma frequency band signaled novelty and graded degrees of familiarity, represented the strength and outcome of associative recall, and finally reflected visual feedback on a trial-by-trial basis. We introduce an attractor-based neural network model that provides a plausible first-order approximation to capture the behavioral and neurophysiological observations. The large-scale data and models enable dissociating and at the same time dynamically tracing the different cognitive components during fast, complex, and natural human memory behaviors.