Today, the neural network is used in many fields.
For example, there is Incremental Learning method for chaotic neural network.
Neurons learn according to its internal state on this learning method.
When the neural network is used for the associative memory, it is more efficient to use incremental learning than correlative learning.
It is conceivable that the causes is difference of methods that set connection weights.
Therefore, this study investigates the reason of efficience of incremental learning focusing on distributions and statistics of connection weights,
comparing them with those from the correlative learning.
As a result, when the learning succeeded, the followings were observed:
- Avarage value is positive number which is close to zero;
- Standard deviation has the appropriate value;
- From kurtosis and skewness, the distribution of the connection weights is similar to normal distribution.
It might be useful to balance positive and negative values on the absolute value in the connection weights, and to prevent the values from becoming large.
