Incremental Classifier and Representation Learning

Abstract

Artificial Neural Networks have revolutionized the field of representation learning over the past decade. It has been shown that by jointly training with multiple objective functions, it is possible to learn general representation that are affective for a multitude of tasks. One key limitation, however, is that simultaneous access to all the data of all the tasks is required during training. Any attempts to learn tasks one at a time, called incremental learning, result in poor performance of the system on older tasks, called catastrophic forgetting. This is unlike humans, who can easily learn new tasks without forgetting about others. Several techniques have been proposed over the past year to solve this problem, however the results are far from ideal. Furthermore, there is no single agreed upon benchmark criteria which makes it difficult to compare existing methods. In this paper, we propose a general framework to compare the prominent existing methods. We analyze their strengths and weaknesses, and investigate how they work in unison. Furthermore, we propose a conditional GAN based rehearsal method, a privacy preserving incremental learning method, and a dynamic threshold moving algorithm. We demonstrate that our proposed methods are effective at solving the problem at hand, and provide promising future directions. We also release a framework implementing ours as well as existing methods to facilitate future research in this direction.