Analyzing Tongue Images Using a Conceptual Alignment Deep Autoencoder

Artificial intelligence can learn some concepts by analyzing sensory data similarly to humans. This paper explores how artificial neural networks (ANNs) can learn abstract concepts by analyzing tongue images based on concepts from traditional Chinese medicine (TCM), which is a discipline that relies heavily on practitioner experience. A computer-aided method will be investigated that analyzes sensory data for TCM practitioners. This paper proposes capitalizing on deep learning techniques. A method called the conceptual alignment deep auto-encoder (CADAE) is proposed to analyze tongue images that represent different body constitution (BC) types, which are the underlying concepts in TCM. In the first step, CADAE encodes the images to a representation space; in the second step, it decodes the patterns. The experiments demonstrate that CADAE can learn effective representations of abstract concepts aligned with BC types by encoding the tongue images. Furthermore, the representation space of the hidden conceptual neurons can be visualized by a decoder network. The experiments also demonstrate that ANNs acquire different data perspectives when different loss functions are used for training. Numerous representation spaces of ANNs remain to be explored. To some extent, our exploration demonstrates that artificial intelligence (AI) has the ability to learn some concepts in a manner similarly to human beings. Based on this ability, AI shows promise in helping humans form new effective concepts that can facilitate medical development and alleviate the burdens of medical practitioners.