This paper proposes a framework for multi-class pattern recognition based on compression of high dimensional features. In pattern recognition, since it is generally difficult to specify features that are suitable for classification of each class in advance, it is important to take the features into consideration broadly from various viewpoints in feature extraction from input pattern. However, if high dimensional features acquired as a result are directly used for classification, various problems such as the explosion of amount of calculation will be caused. Then, high dimensional features in which many features considered to be useful in classification were included are effectively compressed by a neural network to low dimensional features. This realizes highly precise and high-speed recognition.