Group3: Human level knowledge and concept acquisition Activities


Group leader
  • Tatsuya HARADA (The University of Tokyo)
    Concept learning and multimodal recognition
Principal investigator
  • Yusuke KUROSE (The University of Tokyo)
    Corpus development for semantic understanding
  • Lin Gu (RIKEN (AIP) )
    Continuous Learning and Memory Mechanism
  • Yusuke MUKUTA (The University of Tokyo)
    Research on time series prediction and estimation of the causality measure
  • Jun SUZUKI (Tohoku University)
    Natural language processing


Our work focuses on human level acquisition of knowledge and concepts by CAs which will be necessary for advanced cognitive abilities of CAs, knowledge sharing among CAs, and understanding of the operator's intentions by Cas.

In order for CAs to be highly autonomous, it is necessary for them to acquire information about the environment and their tasks in a manner similar to that of humans. For this level of acquisition to be possible, we require technology which combines visual information with human-level knowledge and concepts and information from various modalities such as emotion and gaze. This will enable CAs to acquire knowledge and concepts that can be shared with humans, and by using these concepts, the CAs will be able to understand intentions of the operator or user. To this end, we will develop a new deep learning method that can learn even with little data, capturing the operator's reactions across remote control.This will enable, on the one hand, the CA to have a cognitive function close to that of a human and to understand the operator's intentions appropriately, and, on the other hand, the operator to naturally interact with the user through the CA with advanced cognitive functions, sharing concepts with the user. We will be working in close collaboration with other Groups.


We will develop technologies which enable CAs to acquire the human level knowledge and concepts necessary for understanding the operator's intentions. The knowledge acquisition function is evaluated by the recognition rate against a standard database that can objectively evaluate the performance based on the knowledge. The concept acquisition function is evaluated by assessing how human-like the concept is by direct comparison with humans’ concepts.