講演情報
[17p-S4_202-10]In-Materio Reservoir Computingfor Systematic Evaluationof WearableHumanActivity RecognitionTask
〇(M2)Tu Truong Huynh1, Muhammad Syahmi Hazim Bin Mohd Niza1, Xu Muzhen1,2, Yuki Usami1,2, Hirofumi Tanaka1,2 (1.LSSE Kyutech, 2.Neumorph center Kyutech)
キーワード：
Echo state networks、In-materio reservoir computing、human activity recognition
In-materio reservoir computing based on nanomaterial dynamics offers a promising route toward low-power and hardware-efficient information processing. In our previous work, we developed a customized echo state network (ESN) framework that accurately emulates the temporal dynamics of physical in-materio reservoirs, enabling reliable algorithm–hardware comparison beyond conventional ESN models. Building on this foundation, the present study evaluates in-materio reservoir computing on a real-world task, human activity recognition using the WEAR inertial dataset, which involves classifying 18 activities from noisy wearable sensor signals. Using the simulation framework, we systematically analyze reservoir utilization strategies and identify an effective approach in which the reservoir functions as a temporal feature expansion module for a lightweight classifier. This strategy achieves competitive performance relative to deep learning models while significantly reducing computational complexity and training overhead. Furthermore, preliminary experiments replacing the simulated reservoir with a physical in-materio reservoir device demonstrate comparable performance, confirming the feasibility of transferring the simulated pipeline to hardware. These results advance task-level benchmarking and highlight the practical potential of nanomaterial-based reservoir computing systems.