Improvement of Predictability by Integrating Deep Learning with Symbol Processing

A01Perception and prediction Improvement of Predictability by Integrating Deep Learning with Symbol Processing

 In recent years, deep learning has attracted much attention Especially, generative models becomes increasingly important in the context of deep learning, and it is presumed that integration of generative models and symbol processing should be crucial from the perspective of development of future deep learning technologies and also for the comparison to human brain. In this research, we focus on the following two points.
First, we try to construct deep Q networks (DQNs) integrated with symbol processing. DQN is famous for its usage in AlphaGo, which is a combination of deep learning with reinforcement learning. We further develop the algorithm so that it can achieve inferences and searches by integrating reinforcement learning with symbol processing through, for example, chunking actions and abstracting state representations.
Second, we try to build technologies to transfer images and sentences. We attempt to realize translation through image generation, and deduction inference based on simulation in image spaces.
Through these studies, we seek what kind of symbol/language processing is needed as well as the perception by deep learning for the advanced intelligent functions. Comparing the findings with those in brain science would enable us to make new algorithms and provide new insights for brain science.


  • Yutaka Matsuo

    Project Leader

    Yutaka Matsuo

    The University of Tokyo



  • Yusuke Iwasawa


    Yusuke Iwasawa

    The University of Tokyo

    Project Assistant Professor

  • Masahiro Suzuki


    Masahiro Suzuki

    The University of Tokyo

    Project Researcher