Multiple temporal credit assignment rules achieve comparable neural data similarity

Yuhan Helena Liu, Guangyu Robert Yang, Christopher J Cueva

Keywords: Computational neuroscience, biologically plausible learning rules, temporal credit assignment, recurrent neural network, neural data analysis, representational similarity analysis, neural representations

TL;DR: This study reveals that models trained with gradient approximation-based biologically plausible rules achieve neural data similarities comparable to those trained with Backpropagation Through Time (BPTT).

Abstract: In the quest to understand how the brain's learning capabilities stem from its ingredients, developing biologically plausible learning rules presents a promising approach. These rules, often relying on gradient approximations, need to be examined for their effectiveness in areas other than task accuracies. This study assesses whether models trained with biologically plausible learning rules can emulate neural data similarity achieved by models trained with Backpropagation Through Time (BPTT). Employing methods such as Procrustes Analysis, we compare well-known neuroscience datasets and discover that models using approximate gradient-based rules show neural data similarities comparable to those trained with BPTT at equal accuracies. Our findings reveal that model architecture and initial conditions have a more pronounced impact on these similarities than the learning rules themselves. Furthermore, our analysis indicates that BPTT-trained models and their biologically plausible counterparts exhibit similar dynamical properties at comparable accuracies. Overall, these results demonstrate the capability of biologically plausible models to not only approximate gradient descent learning in terms of task performance but also emulate its ability to capture neural activity patterns.

Submission Number: 20