Research team from Tianjin University led by Professor Qinghua Hu made a breakthrough on Deep Learning in the field of Artificial Intelligence (AI) and proposes a comprehensive multi-view learning framework for the first time, which is expected to improve the limitations of Deep Learning and create a Smart Brain with Early Integration and Analytical Thinking.

These work has been published on IEEE Transactions on Pattern Analysis and Machine Intelligence （IEEE TPAMI）.

The significance of AI lies in the liberation of human labor and the intellectualization of machines. Deep learning is a computational method that makes machines more intelligent. The key is to perform representational learning on images, sounds, and texts, and to interpret how these data imitate human brain mechanism. At present, the mainstream deep learning algorithm is “not very smart”, and there are defects such as single-view analysis and difficulty in obtaining regular cognition. How to combine complex multi-source information for data analysis? How to keep the machine “eyes and ears all open” to all kinds of information and then analyze and think comprehensively? This is a huge challenge for deep learning algorithm research.

The research team led by Prof. Qinghua Hu took the lead in developing the Comprehensive Multi-view Learning Framework Algorithm, and innovatively proposed the research idea of " early integration for multi-source information, joint learning with specific tasks". Compared with the previous Deep Learning algorithm, the innovation of this new algorithm mainly lies in the following two aspects: First, the cross-modality and cross-dimensionality information "early integration" is realized, and the big data in different fields are summarized into low-dimensional comprehensive representation. The second is to build a mathematical model that allows the machine to learn consciously, instead of "stacking analysis" of large amounts of data, and this is realized through a reasonable analysis of the comprehensive network data in order to obtain a principal analysis. It is even possible to predict and infer complex tasks, which is expected to realize a leap from traditional "Deep Learning" to "Integrated Thinking".

At present, this work has been successfully applied to the prediction of infant brain development and the diagnosis of Alzheimer's disease.

By: Li Na; Zhang Changqing

Editor: Qin Mian and Keith Harrington