Class-Incremental Learning Using Enhanced Replay and Auxiliary Network

Abstract

The ability of artificially intelligent agents to continuously integrate new knowledge while preserving existing knowledge is an open challenge in artificial intelligence. This is mostly due to the drastic forgetting of existing knowledge when new knowledge is gained, a phenomenon known as catastrophic forgetting. Initial approaches focussed on scenarios where additional information about the given task is available at test time. Recently, there have been attempts to solve this problem in the more challenging class-incremental setting, where the agent must be able to predict out of all the classes seen without any information about the task. Class-incremental problems are considered difficult to solve without access to stored subset of samples from old classes. For this reason, existing algorithms usually rely on the availability of such samples. For practical applications, it is not always feasible to store all the samples from all seen classes. Therefore, it is important to carefully pick those samples that mitigate catastrophic forgetting the most. In this research, we focus on a class-incremental learning problem in image recognition. Although plain random selection strategy has been shown to pick samples that are effective in mitigating catastrophic forgetting, it is not clear if this strategy is still effective when data consists of noisy images. Images can become noisy due to the use of computer graphics such as in medical imaging or due to faulty sensors, low-light conditions or other environmental factors such as in satellite and CCTV images. We propose our algorithm, classifier outputs for the ranking of examples - high (CORE-high) which provides great improvements over random selection when pixel noise is present in the training images. Next, we address the problem of multi-class imbalance in the context of class-incremental learning in image recognition. This occurs because the training set consists of many samples from new classes and comparatively lower number of samples from old classes. As a result, the network is biased towards new classes, which contributes to catastrophic forgetting. To alleviate the bias from the imbalanced dataset, we train an auxiliary network using equal number of samples from both old and new classes while freezing the feature extraction network. The feature extractor and the auxiliary network are used to make predictions on unseen data. Our method, bias correction using auxiliary network (BiC-AN) achieves a superior performance when compared with IL2M, a state-of-the-art bias correction method. The results are promising with minimal computational costs. Furthermore, BiC-AN is flexible and can be combined with existing class-incremental algorithms.