A Self-organizing Classification Algorithm Based On Deep Learning

During studying the formation of language, we found that brain can receive different signal source. Besides, brain can extract the common features from different expressions of things by means of association, pattern recognition and other operations. Consequently, it classifies things by self-organizing and forms an associative memory. Pavlov’s experiments show we have the ability of conditioned reflex. When we see some associated things we can make a quick associative reaction. For example, we will slobber when we hearing some delicacy.The Artificial Neural Network is a very important domain in the field of Artificial Intelligence what suffering from shallow learning to deep learning. The Back-propagation algorithm makes machine learning achieve great success before 2006. We can dig statistical rules from a large number of training samples by the Artificial Neural Network to predict and classify. However, the Artificial Neural Network has fallen into a trough due to the difficulty of the theoretical analysis and the complexity of training process. With the article in the "science" by Professor Geoffrey Hinton in 2006, Deep Learning is beginning to have a huge impact on the academic and industrial sector.Using Restricted Boltzmann Machine (RBM), Deep Belief Network (DBN) and Self-Organization Mapping Net (SOM), in this study, we establish a multi-channel signal association generated model to simulate the process mentioned above. Firstly, DBN is used to extract features from multi-channel signal source, followed by using RBM to coalesce features, and produce a common feature. Finally, SOM algorithm served as classifying things of same type.In the process of language learning and writing, we can flexibly convert different languages and different text through establishing contact with words in different languages. Previously, we can identify text from different signal input sources by shallow learning without contacting. The paper verifies the feasibility of this algorithm by implementing both the mutual generation between the Arabic numerals picture and Chinese numerals picture, and the self-organizing classification of concepts.We can generate a replaceable expression representing the same kinds of things, which is a conditioned reflex process actually, by the self-organizing classification algorithm based on deep learning through theoretical analysis and empirical analysis. Meanwhile, we succeed learning a unity concept, which realizing the purpose of generating concept.