A Neuron Model Capable of Learning Expansion/Contraction Movement Detection without Teacher's Signal

機関リポジトリ HANDLE

この論文をさがす

抄録

Neuron has the characteristic of reacting to a speci c stimulus. The char- acteristic is said to be from the dendritic morphology of neuron. A neuron which reacts to a speci c stimulus has its unique dendritic morphology. Traditional McClloch-Pitts neuron model failed to include such dendritic functions. In this paper, we propose a neu- ron model that includes such nonlinear functions on dendrite and show that the model is capable of learning Expansion/Contraction movement detection without teacher's sig- nals. The proposed model consists of the retina, LGN (lateral geniculate nucleus), V1 (primary visual cortex) and MST (medial superior temporal area). The neuron model of MST learns the Expansion/Contraction movement detection function by plasticity. Plas- ticity of the model neuron is expressed by back-propagation-like algorithm. Furthermore, we propose a method of creating teacher's signals automatically from the output state of the neuron in MST. We initialize the model neuron with an arbitrarily dendrite randomly and use the model neuron to learn to detect the movement of Expansion/Contraction. Our simulation results show that the model neuron can learn the movement detection of Expanision/Contraction pattern without teacher's signals and can develop its dendritic structure, such as the location of synapses and type of synaptic inputs by eliminating un-useful dendritic branches and synapse.

金沢大学理工研究域電子情報学系

収録刊行物

詳細情報 詳細情報について

問題の指摘

ページトップへ