|Affiliation||Department of Computer Science and Engineering|
|Fields of Research||Computational Intelligence / Neural Informaion Science|
|Degree||Dr. of Engineering (Chiba University)|
|Academic Societies||Institute of Electronics, Information and Communication Engineers (IEICE) / Japanese Neural Network Society (JNNS) / Japanese Congnitice Science Society (JCSS) / Vison Society of Japan / Japan Neuroscience Society / Society for Neuroscience|
Please append "tut.jp" to the end of the address above.
|Laboratory website URL||http://www.ci.cs.tut.ac.jp/|
|Researcher information URL（researchmap）||Researcher information|
Theme1：Neural Circuit Model forming Semantic Network
It is said that humans store the semantic network, which is a linked network of knowledge. We suggest a neural circuit model explaining how a human being learns a network of such a semantic network. The outline of the model is shown in Figure. the association cortex 1 layer passes the inputs to the dentatus gyrus, the CA1, and the association cortex 2 layers, and the CA1 layer produces the learned time sequence patters from the association cortex 1 layer via the dentatus gyrus and the CA3. The association area 2 layers are consists of two layer: 2a and 2b; 2a stores episodic memory as personal experience and 2b stores semantic memory. STDP (Spike-Timing-Dependent synaptic Plasticity: STDP) was used to learning rule in association area 2b to examine the relation of STDP observed in physiology and forming semantic memory. STDP is the phenomenon that transmission efficiency changes by relative timing of the firings of connected neurons, and it is thought as the origin of various learning in the living thing. The thick arrows of figure 2 show the nerve connections that change by STDP.
Theme2：Flexible reinforcement learning algorithm
Humans learn from the result (reward) that they obtained by doing action through try and error, and can perform reinforcement learning to acquire the best action. Such reinforcement learning had been suggested, however, the conventional learning method had the problem of taking much time in re-learning. Therefore, we have proposed a proper parameter control method in reinforcement learning, that is, flexible learning method. In figure 3, the new wall is newly located on the learned path from a start on the left-up to the goal of the right-down. Using the conventional reinforcement learning algorithm, the detour was not able to be searched out easily. Compared with our proposed technique, it was able to find the detour quickly.