概要:

This study aimed to determine the proper amount of weight change ($\Delta w$) for dynamical recall. (Dynamical recall is a distinctive phenomenon on chaotic neural networks. $\Delta w$ is a parameter used for learning patterns.)

Some experiments were conducted to examine how many patterns the network outputs in dynamical recall. In these experiments, the network size was changed from 50 to 600 in increments of 50, and $\Delta w$ was changed from 0.00625 to 0.05 in increments of 0.00625.

These experiments showed the following results.

1. On a small size network (less than 400 elements), proper $\Delta w$ for dynamical recall depends on learning patterns.

2. On a large size network (more than 400 elements), proper $\Delta w$ is approximately constant (0.01875) regardless of learning patterns or the network size.

According to the first result, it can be said that a network with a certain number of elements should be used when conducting experiments on dynamical recall. Otherwise the performance of dynamical recall can not be evaluated accurately.

As for the second result, this study does not show the reason why proper $\Delta w$ is 0.01875.