Search Papers | Poster Sessions | All Posters
Poster A26 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Geometry of naturalistic object representations in models of working memory
Xiaoxuan Lei1 (), Takuya Ito2, Pouya Bashivan1; 1McGill University, 2T.J. Watson Research Center, IBM Research
Working memory (WM) is a central cognitive ability crucial for intelligent decision-making. Recent experimental and computational work studying WM has primarily been carried out using categorical stimuli, rather than ecologically-valid, multidimensional naturalistic inputs. Moreover, such studies have primarily evaluated WM on single or limited numbers of tasks. As a result, there is a lack of understanding in how naturalistic object information is processed by neural circuits. To bridge this gap, we developed sensory-cognitive models, consisting of a convolutional neural network (CNN) coupled with a recurrent neural network (RNN), and trained them on nine distinct N-back tasks using naturalistic stimuli. By examining the RNN’s latent space, we found that: 1) multi-task RNNs simultaneously represent both task-relevant and irrelevant information while performing tasks; 2) the latent subspaces used to maintain specific object properties are largely stable across tasks in vanilla RNNs but not in gated ones; and 3) RNNs embed objects in new representational spaces in which individual object feature axes are more orthogonalized compared to the perceptual space, enhancing separation of features. Our findings elucidate the ways in which goal-driven RNNs adapt their latent representations in response to task requirements.
Keywords: working memory recurrent neural networks multi-tasking representation analysis