Continuous learning unified framework based on deep neural network

Abstract

The invention provides a continuous learning unified framework based on a deep neural network. The continuous learning unified framework comprises the following steps: gradually learning a plurality of tasks through the weight of the deep neural network; wherein the plurality of learning tasks are learned through non-forgetting, forward transmission, confusion prevention and backward transfer; andcombining the weights of the learning tasks, and setting a hyper-parameter b value. According to the framework, all attributes can be demonstrated by using a small amount of weight merging parametersin the deep neural network through the framework; besides, the behavior and mechanism of the framework are similar to those of human learning, such as non-forgetting and forward transmission, so thatconfusion and backward transfer are avoided. As a bidirectional inspiration channel, continuous learning of the machine and human beings is further understood.

Description

technical field [0001] The invention belongs to the technical field of machine learning, and in particular relates to a unified framework for continuous learning based on a deep neural network. Background technique [0002] Humans have the ability to continuously acquire knowledge and skills, refine them based on new experience, and transfer them across domains throughout their lifespan. Human learning abilities have been inspiring machine learning methods for decades, and the goal of this work is to further expand the impact of human learning on continuous or lifelong learning in machines. At present, the method of machine learning is usually to use all the neural networks from the beginning, so there are some problems. For example, it is necessary to learn a difficult task through the deep neural network, or to learn more than two tasks. Start learning from difficult tasks or learn more than two tasks at the same time, so that it will be very difficult to learn, there are...

AI Technical Summary

Problems solved by technology

At present, the method of machine learning is usually to use all the neural networks from the beginning, so there are some problems. For example, it is necessary to learn a difficult task through the deep neural network, or to learn more than two tasks. Start learning from difficult tasks or learn more than two tasks at the same time, so that it will be very difficult to learn, there are too many things to learn, and the required learning time is longer; A task, after a period of time, it is found that the second task needs to be learned, and the weight of the deep neural network is not enough, and it may not know, then it will probably take the learned first task when learning the second task. task coverage

Images