This project aims to develop an AI system that can recognize abstract concepts defined as “deep Shitsukan,” as humans do. With the development of deep learning, AI technologies have made dramatic progress in recent years, but the current AI is merely “cognitive automation” (i.e., the automation of human cognitive processes). The same approach cannot achieve the above goal because Shitsukan is difficult to verbalize and quantify and thus almost impossible to annotate for supervised learning. Besides, since Shitsukan is context-dependent, we cannot extract Shitsukan recognition as a stand-alone task like object recognition. We must view it as a problem of comprehensive image understanding. Toward the above goal, we hypothesize that a neural network that can understand images like humans will acquire feature representation of various Shitsukan representations. We will first improve image understanding tasks and then establish a method for extracting representations inside the network, aiming to achieve the above goal.

Co-Investigator