Incremental Learning Based On Bayesian Optimization And Resource Release Mechanism

Humans and animals can continuously accumulate and consolidate knowledge gained from previously observed tasks and constantly learn to solve new problems or tasks.However,such continuous learning capabilities are one of the major difficulties for machine learning systems,as the constant acquisition of available information from non-stationary data distributions often leads to catastrophic forgetting.With the development of Deep Learning in recent years,many ways to overcome catastrophic forgetting have emerged,but continuous learning remains a long-standing challenge in Artificial Intelligence.Aiming at some problems still existing in the field of continual learning,this paper combines Bayesian optimization and variational inference to propose a continual learning algorithm based on uncertainty and resource releasing mechanism and Bayesian continuous learning which promotes the regularization operation to the level of neurons.The main research work is as follows.In order to carry out Bayesian inference more efficiently,a Bayesian model is optimized by minimizing the KL divergence between the true posterior distribution and the prior distribution with the Gaussian mixture distribution as a priori and the variational inference.In order to adapt the model to incremental scenarios,the importance judgment mechanism of parameters is introduced,and the learning rate of parameters is dynamically adjusted according to the uncertainty(standard deviation of probability distribution)carried by Bayesian neural network,so as to avoid over-adjustment of parameters with high importance(low uncertainty).Then,in order to promote the learning efficiency of future tasks,a resource releasing mechanism is proposed,which provides more space for the learning of subsequent tasks by selectively abandoning part of the resources from the old tasks through the guidance model.Using Monte Carlo sampling to solve the true posterior of the Bayesian model.Although the true posterior of the model can be approximated more accurately by Monte Carlo sampling,the operation time of the model is too long and the efficiency is so low.Secondly,the first work is still at the weight-level,that is,to impose different intensity restrictions on the weights according to the importance of the individual weights,which will lead to the conflict of information flowing to the neurons.To solve these two problems,first of all,all the variables of the model are assumed to be independent,and then the KL divergence term is analytically computed by factoring the term into multiple individual Gaussians,which saves more calculation cost.Secondly,the input weights are grouped by neurons and the variance of the weights is restricted to the same value according to the group,and then the unified variance is used as the importance discrimination basis to regularize the weights.The performance evaluation on public datasets MNIST,Fashion-MNIST and CIFAR10 shows that the proposed method maintains the highest classification accuracy in several neural networks with different sizes,and the classification performance of the proposed method also maintains the minimum decline when the network size is reduced.