Title: A differentiable point process with its application to spiking neural networks (ICML-21)
A spiking neural networks (SNN) is an artificial neural network (ANN) where neurons communicate with each other using spikes rather than real values as the conventional ANNs do. While SNN is expected to be a more energy-efficient alternative to the conventional ANNs, a lack of standard learning algorithms discourages us from its real-world applications as well as basic research on it.
In this talk, we start by introducing SNNs including its probabilistic formulation using point processes. Then, we review the existing learning algorithm by Jimenez Rezende & Gerstner (2014), and finally, present our learning algorithm, which uses a differentiable point process as a building block. We investigate the effectiveness of our learning algorithm through numerical simulation.
Hiroshi Kajino is a research scientist at IBM Research – Tokyo. He obtained a PhD degree in Information Science and Technology from the University of Tokyo in 2016, and joined IBM in the same year. His current research interest focuses on machine learning for molecules and spiking neural networks.