Kiyohiko NAKAMURA Laboratory Intelligence and The Brain Link to Japanese Page

About Nakamura's Lab

Research areas

Brain science, Computational neuroscience, Neurophysiology, Machine intelligence

Research methodology

Our goal is to know how the brain works and to show how to design high-level machine intelligence based on the knowledge. Theoretical approach: We collect neuro-physiological and -anatomical data on certain brain regions and build mathematical models of their neural circuits. We analyze the biologically realistic models to see how the circuits perform their functions, which are impaired by damage of those regions. If you will build biologically realistic models of more regions of the brain and connect them in the same way as in the real brain, we believe that the integrated model will implement higher-level intelligence of the brain. Experimental approach: First, we predict change in states of components of the model. To test hypotheses described by the models, we conduct electrophysiological experiments to see whether the brain includes neural activity parallel to the predicted change.

Research issues

  1. Neural mechanism of information measurement.

    Animals seek information to reduce their efforts to receive rewards and perform actions that enable them to gain more information. The ability of seeking information serves higher cognitive processes such as planning and reasoning. This study presents results indicating that the brain quantifies information by using the information-theoretic measure: In the premotor cortex of the monkey required to choose the most informative target from multiple alternatives, we found neural activity encoding information value given by Shannon information measure.

    Most informative test

    Ref.: Kiyohiko Nakamura (2006) Neural representation of information measure in the primate premotor cortex. Journal of Neurophysiology 96: 478–485.

  2. Multistable attractors in a network of phase oscillators

    Neurophysiological experiments have shown that some prefrontal neurons keep their level of activity for several seconds to encode an analog quantity in the firing rate of multistable neuronal networks. We suggested the possibility that working memory uses the degree of synchrony among neurons to encode an analog quantity.

    Ref.: Takuma Tanaka and Toshio Aoyagi (2011) Optimal weighted networks of phase oscillators for synchronization. Physical Review Letters, 106, 224101.

  3. Neural processing in the subsecond time range in the temporal cortex of the cerebrum.

    Primates recognize and respond to complex stimuli within half a second. This study presents a hypothesis that cortical processing of the millisecond time range is performed by latency competition between neuronal populations and tests it with a biologically realistic model of the monkey temporal cortex. The model is a neural circuit of cortical areas, V1, V2/V3, V4, PIT, CIT, AIT, and STPa. Each area is an array of neuronal populations which inhibit each other. The model predicted neural activity recorded at the monkey STPa.

    Ref.: Kiyohiko Nakamura (1998) Neural processing in the subsecond time range in the temporal cortex, Neural Computation, 10(3): 567–595.

  4. Neural mechanism of reinforcement learning in the neocortico-thalamo-hypothalamic system.

    Animals learn how to survive in unfamiliar environments by trial and error. This study presents hypothetical mechanisms of the neocortico-thalamo-hypothalamic system which associates sensory stimuli with rewarding actions. Learning by trial and error requires two functions: (1) generating various trials and (2) finding rewarding actions out of the trials according to rewards. The present model suggests these are implemented by the neocortico-thalamic system and the the neocortico-hypothalamic system respectively.

    Ref.: Kiyohiko Nakamura (1993) A Theory of Cerebral Learning Regulated by the Reward System. I: Hypotheses and Mathematical Description, Biological Cybernetics, 68(6): 491–498.

  5. Neural circuits for probabilistic reasoning.

Future research

We are interested in concept learning in which cognitive ambiguity about the concept is reduced by acquired external information. We quantify the ambiguity using information entropy and model the learning process as combination of generalization according to similarity and maximization of expected Shannon information in the external information. To test the model is our goal of current research.

Ref.: K. Nakamura, A. P. Sage, and S. Iwai (1983) An intelligent data-base interface using psychological similarity between data. IEEE Trans. on Systems, Man, and Cybernetics, 13: 558–568.


Please refer to this list.

4259 Nagatsuta-cho, Midori-ku, Yokohama 226-8502 JAPAN
Last Update 04/11/14