Research On Continual Learning Classification Technology Based On Experience Replay

In the era of big data,with the revival of deep learning and the continuous ex-ploration of algorithms based on deep learning in the fields of machine learning and computer vision,artificial intelligence has made great breakthroughs in the past decade,and its performance can even be comparable to that of human beings in some fields.However,when extended to practical scenarios,these agents usually designed for a sin-gle task do not have robust adaptive ability.Once they learn new knowledge,they will greatly ”catastrophic forget” the old learning information,which hinders their further development to some extent.In order to solve this bottleneck,researchers proposed Continual Learning(CL)technology to try to alleviate the problem of catastrophic for-getting in deep neural networks.This technology simulates the human brain’s behavior habit of learning knowledge,and constructs the experience replay in the rehearsal pro-cess with the help of the instances of old classes stored in the memory buffer,so as to consolidate the old knowledge while learning new knowledge.By constructing differ-ent experience replay methods,this thesis proposes an algorithm framework based on two kinds of rehearsal to realize continual learning.Aiming at the assumption that the performance of replay-based methods depends on the rehearsal mode and the effectiveness of replay exemplars,a Coordinating Expe-rience Replay(CER)algorithm is proposed in this thesis.The algorithm excavates the commonness and individuality between rehearsal modes through harmonious experi-ence,organically integrates direct rehearsal and indirect rehearsal,and supplements the missing part through Imitative Experience Replay,so as to realize the mutual promotion between multiple rehearsal modes.In addition,in order to better realize the selection of replay exemplars and ensure that replay exemplars can characterize the characteristics of their classes while taking into account the balance between classes,Vital Exemplar Reserved Sampling strategy is proposed to provide more appropriate replay exemplars for the rehearsal process,and then better cooperate with a variety of rehearsal modes.From the perspective that the relationship between instances is more representative than individual information,based on the assumption that self distillation learning can better retain knowledge,this thesis proposes Correlation-consistent Experience Replay(C-CER)algorithm.The algorithm constructs a rehearsal method with richer compar-ison and similarity relationship between instances in the replay process.Among them,the comparison correlation consistency matching module maintains the consistency of the distribution of triples,and the similarity matrix between different instances is cal-culated by the similarity correlation consistency matching module to replace the single constraint based on the instance’s own information.Furthermore,this scheme constructs the self distillation learning mode of the past time and the current time model,forms the consistency replay of the above multiple associations,so as to retain the learned knowl-edge and optimize the learning process of the model.In this thesis,extensive experiments are conducted on seven mainstream continual learning image classification datasets for Traditional Continuous Learning(TCL)and General Continuous Learning(GCL)to verify the effectiveness of our proposed meth-ods.At the same time,this thesis designs experiments to analyze each module of the proposed method in detail,and explores the internal mechanism of some modules.