| 研究実績の概要 |
The principal investigator has recently changed the direction of the study, because the principal investigator has found that the old learning rule has some problems. He found that the learning does not converge to meaningful receptive field. In particular, the units in the middle layer biased to some of the patterns, even though the training order is randomized. Also, not all realization with different initial states lead to qualitatively similar final outcome. So the principal investigator believed that the model with that settings does not deserve resources for further investigation.
In order to solve this problem, the principal investigator has inputted a new learning rule and study their properties. Recently, I found that the middle layer with a dentate-gyrus-like learning rule can separate input patterns without supervising signal, which deserve further studies. Also, the principal investigator found that the output of the middle layer is very helpful to pair with specific labels. For example, the investigator has used MNIST to test the network’s ability. He found that the correct rate can be up to 95% with only one learning iteration. Also, the current learning rule could generate unstable network activity, although the tuning is helpful for pattern separation, which is also a interesting point to follow.
|
