Deep neural network incremental training method and system based on a learning automaton

Abstract

The invention provides a deep neural network incremental training method and system based on a learning automaton. The deep neural network incremental training method comprises the steps of an original sample set training step: training a deep neural network model by adopting an original training sample set; and an incremental sample set training step: carrying out category expansion on the previous training sample set, and carrying out incremental training on the previous trained deep neural network model by adopting the expanded training sample set. According to the method, the training process on the complex data set can be decomposed into a series of simple task training processes, and the classification problem of all categories is solved step by step through model incremental training. The cumulative learning strategy better conforms to the learning process of people in real life.

Description

technical field [0001] The invention relates to the field of information processing, in particular to a learning automaton-based incremental training method and system for deep neural networks. Background technique [0002] Neural network is a machine learning algorithm designed to simulate the hierarchical connection structure of biological brain neurons. It can learn potential feature transformations from sample data and realize nonlinear mapping from input to output. In recent years, with the rapid development of big data and high-performance parallel computing equipment (such as general-purpose graphics processor GUGPU), deep neural network technology has been extensively studied and achieved remarkable results, such as image classification, object detection, speech recognition and even Superhuman performance continues to be achieved in fields such as board games. [0003] The basic unit of a neural network is a neuron, which consists of a specific number of neurons to ...

AI Technical Summary

Problems solved by technology

If the model trained on a certain sample set continues to train on new samples, it will face the problem of catastrophic forgetting, that is, after training on new samples, the original features in the model will be quickly destroyed, which is equivalent to re-training on new samples. The entire model is retrained on the expanded sample set, which cannot make good use of the previously learned features

Therefore, the current deep neural network training algorithm does not have the ability to perform incremental and continuous learning when the training data gradually increases.

Images