[1Win4] Poster session 1

[1Win4-10] Utilization of network with Bayesian inference unit to predict input confidence

〇Ayumu Yamada¹, Takashi Takekawa¹ (1.Informatics Program, Graduate School of Engineering, Kogakuin University of Technology and Engineering)

Keywords:Bayesian Inference, Computational Neuroscience, Hopfield network

The free energy principle suggests that the human brain may perform probabilistic inference based on Bayesian inference. In this study, we propose a new Bayesian inference unit that focuses on the ability of probabilistic inference to handle ambiguity, which is a characteristic lacking in deterministic inference such as deep learning. This unit aims to reduce computational load by predicting the confidence of new inputs from past learning and preemptively excluding inputs with low confidence. Furthermore, considering this as a fundamental characteristic of the brain, we investigated the behavior of a Hopfield neural network (HNN) applied with the proposed unit as neurons. The experiments demonstrated that the proposed HNN can recall embedded patterns or exhibit unstable states.