Presentation Information
[2Yin-B-41]Multi-timescale homeostasis based on energy metabolic constraintsAutomatic Acquisition Mechanism of Stable Spontaneous Firing in Self-Organizing Spiking Neural Networks
〇Atsushi Katayama1, Kazuhiko Murasaki1, Ryuichi Tanida1 (1. NTT)
Keywords:
SNN,homeostasis,energy finiteness,automatic adjustment
Homeostasis in Spiking Neural Networks (SNNs) is a vital mechanism for balancing functional stability and information processing efficiency. This study introduces a biologically inspired approach to automate parameter tuning by focusing on the physical constraints of metabolic energy, such as ATP. By defining "energy finiteness" as a core cost function, each neuron monitors its energy consumption. Based on a preset energy budget, the system dynamically regulates the intensity of homeostatic mechanisms across different time scales: Spike Frequency Adaptation (SFA), Intrinsic Plasticity (IP), and Synaptic Scaling (SS). This allows for the automatic derivation of parameters based on the energy balance per unit network, ensuring scalability. This framework integrates SFA to reduce redundant energy use, IP to maximize information transfer, and SS to maintain long-term metabolic stability. Collectively, these mechanisms provide a robust foundation for low-power, high-level computation, mimicking the metabolic efficiency of the biological brain.