主催: 一般社団法人 日本機械学会
会議名: 第35回 計算力学講演会
開催日: 2022/11/16 - 2022/11/18
Deep learning has problems such as "inference is a black box", "unexpected answer by overfitting", and "large-scale network and long-time learning". Bayesian inference performs learning and inference that is completely different from neural networks There are two biggest differences between neural networks and Bayesian inference. The former is a data-dependent type and the latter is a deductive type. The other is the difference in degrees of freedom, the former is infinite degrees of freedom and the latter is limited degrees of freedom. Therefore, the former is superior in learning ability and the latter is superior in learning efficiency. In this paper, we discuss pattern recognition based on Bayesian inference using MNIST handwritten digit pattern data.