Search Papers | Poster Sessions | All Posters
Poster A23 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Emergence of Modular Structure in the Recurrent Neural Network During Incremental Multi-task Learning
Yuhang Wu1, Shi Gu1; 1University of Electronic Science and Technology of China, Chengdu, China
Biological brain networks universally exhibit distinct simplicity: they are composed of modular components that may function relatively independently. Yet, there's currently no consensus on the origins of modularization. In this study, we trained single recurrent neural networks on multiple cognitive tasks requiring working memory, decision-making, classification, and inhibitory control, thereby simulating real-world challenges. Our findings reveal that under conditions of constrained network size, multitasking promotes greater modularity compared to scenarios involving fewer tasks. This implies that modularity may arise as an adaptation in models required to handle multiple tasks when the units available for computation are limited. Additionally, we compared the learning processes of dynamically evolving networks, which form new connections periodically, with those of statically fixed networks, where connections are pre-established at the start of training. We found that models growing sequentially lead to higher modular structures across all wiring rules. Our study proposes that functional demands consequently influence structural formation, offering new insights into neuroscience.
Keywords: modular neural network computational neuroscience structural-functional interaction cognitive multitasking