[1O3-GS-8] Soft Computing

[1O3-GS-8-01] A Study on Building a Deep Generative Model for Prediction in the Human Brain

〇Eri Kuroda¹, Ichiro Kobayashi¹ (1. Ochanomizu University)

Keywords:Prediction Model, Predictive Coding, Partially Observable Markov Decision Process

In recent years, researches on machine learning and deep learning have been actively studied. Among them, the development of deep generative models for prediction has been attracting attention. In our study, we focus on predictive coding, which is the assumption for prediction ability of the human brain, and attempt to develop a human brain-like prediction model. PredNet is a deep learning model imitating predictive coding, however, it can predict next image with the same sequential interval, but not with any intervals. On the other hand, temporal differential variational autoencoder (TD-VAE) can predict the next image with any sequential intervals but is not a model reflecting human brain function. So, we develop a new human brain-like prediction model by unifying PredNet and TD-VAE so that the model should have the function of predictive coding and can predict with any sequential intervals. Through the experiments with two kinds of datasets, we have confirmed that our proposed model can predict next images correctly.