Since a clustering process can be regarded as a map of data to cluster labels, it should be natural to employ a deep learning technique, especially a feed-forward neural network, to realize the clustering method. In this study, we discussed a novel clustering method realized only by a feed-forward neural network. Unlike self-organizing maps and growing neural gas networks, the proposed method is compatible with deep learning neural networks. The proposed method has three parts: A map of records to clusters (encoder), a map of clusters to their exemplars (decoder), and a loss function to measure positional closeness between the records and the exemplars. In order to accelerate clustering performance, we proposed an improved activation function at the encoder, which migrates a soft-max function to a max function continuously. Though most of the clustering methods require the number of clusters in advance, the proposed method naturally provides the number of clusters as the number of unique one-hot vectors obtained as a result. We also discussed the existence of local minima of the loss function and their relationship to clusters.