Building deep architectures for traditional chinese medicine diagnosis

Traditional Chinese medicine (TCM), which is built on a foundation of a long history of Chinese medical practice, provides a quite different approach with western medicine. It has formed a deep knowledge of medical science, theory, diagnostic methods, prescriptions and cures. Fully grasping the deep knowledge that mostly based on a great amount of experience is beyond anyone's capability. May we train a big electronic brain to assist diagnosis and healthcare suggestion from TCM approach? Thanks to the advances in deep learning that provides us a powerful method for fusing huge volume data into effective representations, we propose a framework of deep architectures that directly maps the heterogeneous high-dimensional sensory data into a common low-dimensional latent feature space, which is quite approaching the analysis of a TCM practitioner. The information process includes images, audios, text and other sensory data. We propose to use a hierarchically supervised training method to train the deep architectures and obtain the interpretable representations. It can facilitate the non-invasive, painless and inexpensive TCM diagnostic methods to check people's health state with a computer assistant.