End-to-End Lifelong Learning: a Framework to Achieve Plasticities of both the Feature and Classifier Constructions

Abstract

Plasticity in our brain offers us promising ability to learn and know the world. Although great successes have been achieved in many fields, few bio-inspired machine learning methods have mimicked this ability. Consequently, when meeting large-scale or time-varying data, these bio-inspired methods are infeasible, due to the reasons that they lack plasticity and need all training data loaded into memory. Furthermore, even the popular deep convolutional neural network (CNN) models have relatively fixed structures and cannot process time varying data well. Through incremental methodologies, this paper aims at exploring an end-to-end lifelong learning framework to achieve plasticities of both the feature and classifier constructions. The proposed model mainly comprises of three parts: Gabor filters followed by max pooling layer offering shift and scale tolerance to input samples, incremental unsupervised feature extraction, and incremental SVM trying to achieve plasticities of both the feature learning and classifier construction. Different from CNN, plasticity in our model has no back propogation (BP) process and does not need huge parameters. Our incremental models, including IncPCANet and IncKmeansNet, have achieved better results than PCANet and KmeansNet on minist and Caltech101 datasets respectively. Meanwhile, IncPCANet and IncKmeansNet show promising plasticity of feature extraction and classifier construction when the distribution of data changes. Lots of experiments have validated the performance of our model and verified a physiological hypothesis that plasticity exists in high level layer better than that in low level layer.