Search Papers | Poster Sessions | All Posters
Poster C25 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Overcoming sensory-memory interference in artificial and biological neural networks
Andrii Zahorodnii1, Diego Mendoza-Halliday1,2, Ning Qian3, Robert Desimone1, Christopher J. Cueva1 (); 1Massachusetts Institute of Technology, 2University of Pittsburgh, 3Columbia University
Memories of recent stimuli are crucial for guiding behavior. However, the same sensory pathways that receive information to be remembered are constantly bombarded by new sensory experiences, and it remains largely unknown how the brain overcomes interference between sensory and memory representations. Here we report which mechanisms might be at play in artificial and biological networks that are robust to sensory-memory interference. We examined recurrent neural networks (RNNs) that were either hand-designed or trained using gradient descent methods, and compared our results with neural data from two macaque experiments. We found an infinite RNN solution space, that included gating of the sensory inputs, modulating synapse strengths to achieve a strong attractor solution, and dynamic coding of feature preference, such that, at the extreme, cells invert their tuning over time. Neural data from macaque brain area medial superior temporal (MST) was most aligned with the Gating + Inversion of Tuning solution. This solution was also consistent with experimental results from monkey behavior. Taken together, our results help elucidate how recurrent neural networks are able to solve the problem of sensory-memory interference using a combination of both static and dynamic codes, and suggest MST may play a role in this computation.
Keywords: working memory attractor recurrent neural networks sensory-memory interference