Dynamical Similarity in Multitasking RNNs: Identifying Shared Motifs Across Task Periods

Biological networks exhibit remarkable flexibility, enabling efficient multitasking. This study investigates the role of dynamical motifs in facilitating this adaptability within recurrent neural networks (RNNs). Instead of focusing on geometric structures, we employ Dynamic Similarity Analysis (DSA) to reveal underlying dynamic patterns. applying DSA to the internal states of RNNs across 15 tasks, we identify the delayed response motif as particularly distinct, characterized by unique rotational properties. Our findings highlight DSA’s ability to uncover differences overlooked by geometric methods, offering deeper insights into the computational mechanisms underlying both biological and artificial intelligence systems.