Study on time-length processing by neural networks

It is thought that 'time-length', which is one of the most important parameters for representing temporal patterns, is coded on some way in the brain. By two methods, we tried to estimate the principle of coding/memorizing time-length, focusing on 'the variation of the encoded time-length image on time'. One was auditory psychophysical experiment on time-length comparisons using the Constant Method. We found that the tendency of the interval of uncertainty, which indicates an amount of forgetting, depended both on first tone length and on interval of two tones. And the other was to evaluate the principle with which recurrent neural networks came to compare two time-length patterns after learning. In these simulations, it was estimated that the time-length information tended to be coded into the position on the trajectories toward the attractors in state space. It was also shown that this network revealed some characteristics of human memory decay when it was fluctuated by random noise.