The present paper aims to propose a method for evaluating and interpreting prototype networks hidden within multi-layered neural networks. The prototype network can be obtained by compressing as many different types of multi-layered networks as possible. However, the final prototype networks are naturally realized by losing information in the original multi-layered neural networks. To preserve the information in the original multi-layered neural networks, we introduce pseudo-learning, where learning essentially involves not learning, thereby retaining the original internal information of multi-layered neural networks as much as possible. Pseudo-learning is implemented by adopting the concept of the conventional learning approach, such as entropy minimization, or in our terms, potentiality minimization. By keeping the weights fixed in prototype networks, we aim to change only the potentiality to simulate a learning process by modifying a parameter for potentiality. The method was preliminarily applied to a simple qualitative bankruptcy dataset. The results confirmed that the compressed networks or prototype networks could be altered with fixed connection weights to evaluate generalization performance. This enabled the interpretation of the internal information transferred from the original multi-layered neural networks.