Presentation Information
[10a-B21-11]Local Information Processing Capacity for Anatomizing Individual Dynamical States
〇(P)Jiaxuan Chen1, Daiki Nishioka1, Yicheng Song4, Ryo Iguchi1, Sota Hikasa1,2, Hina Kitano1,2, Hirose Akira3, Takashi Tsuchiya1,2 (1.NIMS, 2.Tokyo Univ. of Sci., 3.The Univ. of Tokyo, 4.NITech)
Keywords:
Physical reservoir computing,Information processing capacity,State pruning
Physical reservoir computing harnesses intrinsic physical dynamics to perform computation, yet the diversity and complexity of these dynamics make them difficult to interpret. While metrics such as information processing capacity (IPC) effectively characterize global computational capability of the overall system, the lack of principled local measures for individual dynamical states obscures internal structure and hinders system optimization. Here, we introduce local information processing capacity (L-IPC), which systematically quantifies individual dynamical states by representing them as vectors in an infinite-dimensional functional space. Each state’s computational properties are then characterized through its projections onto subspaces of specific nonlinearity orders and memory depths, while its residual norm captures computational irrelevance such as noise. We validate L-IPC on echo state networks and then apply L-IPC to two physical reservoirs based on ion-gating and spin-wave dynamics. We reveal their internal computational structure and demonstrate task-adaptive state pruning, substantially outperforming conventional pruning strategies. These results establish L-IPC as a practical and principled tool for interpreting and optimizing neuromorphic computing systems.