|Date and Time:
||August 4 (Fri.), 2017, 13：00 – 14:30
||Meeting room #306, Building E-3, UEC
||Dr. Tomoki Fukai (Laboratory Head, Laboratory for Neural Circuit Theory, Brain Science Institute, RIKEN)
||Prof. Shigeru Tanaka
||Brain’s network mechanisms to model the external world
||How does the brain model the external world? What circuit mechanisms underlie this computation? These questions are deeply connected to the biological mechanisms of learning. To understand the underlying mechanisms of brain’s modeling of the external world, I show two recent computational studies from my group. The first topic is sequence learning in the hippocampus. Recently, the role of spontaneous activity, or preplay, in the formation of place-cell sequences has been debated. Preplay is a phenomenon in which spontaneous sequences preexisting before experience turn into place-cell sequences during spatial navigation. Preplay suggests that the innate structure of cortical circuits is useful for information coding, and hence is conceptually interesting. We construct a recurrent network of two-compartment neurons to utilize spontaneous sequences for robust one-shot learning of place-cell sequences. Our model suggests that dendritic computation in pyramidal cells is crucial for this processing. The second topics is chunking, which is the ability of the brain to detect repeated segments of patterns from complex sequences. I will present a novel mechanism of chunk learning in recurrent neural networks. The key is to use independent multiple reservoir computing systems that supervise each other during learning. Interestingly, readout neurons in our model behaves like “stop cells” which have been discovered in the basal ganglia after motor sequence learning.