Shigeo Sato

The paper investigates how input encoding and spike detection methods affect neural representations and reservoir computing performance in biological neural networks for multiclass time-series classification.

This study uses microfluidic platforms to investigate how damage to hub or peripheral connections in modular neuronal networks affects recovery, finding that hub damage leads to delayed recovery through alternative pathway formation.

This paper develops a closed-loop biological neural network system that learns and generates temporal patterns through feedback control, demonstrating adaptability and reduced synchronization.

This paper introduces Predictive E-prop, a biologically inspired model that trains recurrent spiking neural networks using gradient-based prediction error minimization, achieving self-supervised learning of complex time-series signals.

This study examines how e-prop learning performs in recurrent spiking neural networks under varying connectivity and noise levels, finding that sparsity improves learning by enhancing activation diversity.