Automatic neural network learning for human behavior understanding

Understanding human behavior is one of the most pivotal steps toward real-world Artificial Intelligence (AI) or even Artificial general intelligence (AGI). However, this task is challenging as human social attributes make human beings unique, leading to various and complicated behaviors. Moreover, real-life behavior data are normally high-dimensional with dynamic changes or even non-Euclidean structures, involving multiple modalities.

Currently, one of the first alternatives to addressing these challenges is using deep neural networks or deep learning, which has brought revolutionary changes in data computation and computer sciences. Nevertheless, expert knowledge of both neural architecture design and human behavior analysis is expected more than ever before in this interdisciplinary research field. All these issues spur the current deep learning studies towards automatic deep neural network learning, which could automatically sketch a neural architecture for a given behavior analysis task.

In line with this topic, this thesis explores the automatic neural network learning approach for human behavior understanding from the most representative behaviors, including human facial expression and actions, step by step. First, manually designed computational models are proposed for human facial expression and actions with dynamic information and graph structures. Based on this, to free humans from the exhausting process, more advanced methods, i.e., automatic neural network learning, are presented. Extensive experiments on benchmark facial expression datasets and action recognition datasets are conducted and comparison results prove the effectiveness of the proposed methods.