Abstract-The present paper aims to propose a new potential learning method to overcome the problem of collective interpretation for interpreting multi-layered neural networks. The potential learning has been introduced to detect important components of neural networks and to train them, taking into account the importance of components. Recently, it has been applied to multilayered neural networks and then the interpretation of input neurons or variables can be possible by collectively treating intermediate layers. However, the collective interpretation for multi-layered neural networks tends to be instable, because the potentialities computed in the pre-training become different from those in the main training. To overcome this problem, we introduce the potential learning with direct potential assimilation. The direct potential assimilation means that the potentiality assimilation is not applied in the phase of pre-training but it is applied directly to training multi-layered neural networks. The new method was applied to the student evaluation data set. Then, it was observed that the selectivity of connection weights could be increased. Then, the input-output potentiality was quite similar to the regression coefficients of logistic regression analysis. Finally, the new method could extract more explicitly inputoutput relations than the regression coefficients by the logistic regression analysis, while improving generalization performance.