Neurons in the primary visual cortex respond selectively to the orientation of edges and their direction of motion. Recently, the arrangement of direction preference was revealed in ferret cortical area 17 and in cat cortical area 18; it was a mosaic-like map. It was found that some iso-orientation domains are subdivided into selective regions for the opposite direction of motion. We used a variation of Kohonen's SOM, which uses projection learning, to form a direction map with an orientation map. This SOM was observed to form opposite direction selectivity sub-domains in iso-orientation domains.
It is known that there exist polynomial-time solutions to the shortest path problem. However, these algorithms are assumed to be run on the processors, and difficult to be applied to the real time control because of the time limitation. This paper proposes an implementation method of the shortest path problem to the real time control using neural network. The proposed method can be implemented by hardware digital logic circuits, and enables the shortest path design to be 4 to 6 digits faster than the conventional algorithms. The proposed method is also easily implemented by LSI.
There are fundamental differences in the memory system between the brain and that of digital computer. The computer stores and retrieves memories guided by an address information, whereas the brain stores memories over the weight space of neural networks through learning and retrieves them according to some dynamical processes. In the course of establishing long-term memory, hippocampus plays an important role, i. e., to make short-term memory of spatially and temporally associated input information. We (Tsukada, et al., 1996) proposed a spatiotemporal learning rule based on the difference in hippocampal long-term potentiation (LTP) induced by various spatio-temporal pattern stimuli. Essential point of our learning rule is that the synaptic weight changes depending on both “spatial coincidence” and “time history” of input pulses. We compared the pattern discriminating ability of this rule with that of Hebbian and its extention rule, through computer simulation on the one layer neural network model with 24 spatio-temporal input patterns and 120 output neurons. It is shown that the proposed rule has the highest ability in separating different spatio-temporal patterns into the synaptic weight space.