Presentation Information
[8p-E217-12]Multi-Input Facial Recognition using In-Materio Device based on Predictive Coding Reservoir Computing
〇(M2)Tu Truong Huynh1, Pritha Roy1, Arie Rachmad Syulistyo1, Ahmet Karacali1, Muzhen Xu1,2, Yuki Usami1,2, Yuichi Katori3, Hakaru Tamukoh1,2, Hirofumi Tanaka1,2 (1.LSSE, Kyushu Inst. Tech (Kyutech), 2.Neumorph Center, Kyutech, 3.Future Univ. Hakodate)
Keywords:
Materials Computation,Reservoir Computing,Feedback System
This study investigates the implementation of Predictive Coding with Reservoir Computing (PCRC) [1] for facial emotion recognition utilizing in-materio devices. At each timestep, PCRC generates a prediction then uses the error against the target to continuously update the readout weights and modify the input signal via the First-Order Reduced and Controlled Error (FORCE) algorithm [1]. Adapting this paradigm into in-materio devices requires both a multi-input hardware architecture and a real-time feedback system. To construct the multi-input framework, Greedy Forward Selection was applied to extract an optimized 8-feature set from time-series 3D facial landmarks within the RAVDESS dataset [2]. This process sequentially appends a feature yielding the highest gain in accuracy during each simulation round, until the target number of features is reached. Simulation of this raw feature extraction demonstrated a progressive accuracy increase, peaking at 67% with 8 features. Evaluating this architecture on a physical in-materio silver nanoparticle (Ag/Ag2S) device featuring a 4-input and 11-readout (4In-11RO) configuration yielded a classification accuracy of 70% using only a linear classifier. To achieve fully online PCRC capabilities, current development focuses on implementing the requisite hardware feedback loop. The pipeline is being upgraded from offline processing to a real-time execution model.
[1] Y. Yonemura & Y. Katori. Front. Comput. Neurosci. 18, 1464603 (2024).
[2] S. R. Livingstone & F. A. Russo. PLOS ONE 13, e0196391 (2018).
[1] Y. Yonemura & Y. Katori. Front. Comput. Neurosci. 18, 1464603 (2024).
[2] S. R. Livingstone & F. A. Russo. PLOS ONE 13, e0196391 (2018).