Abstract

Artificial intelligence empowered by neural networks (NNs) has tremendously improved the state-of-the-art results in various domains of applications over the past years. Inspired by human intelligence, there has been a growing interest in the real-world learning settings that pose multiple challenges to the learning agents. Continual learning (CL) is a learning paradigm in which the agent learns new tasks continually from a never-ending and non-stationary stream of data. A CL agent should be plastic enough to learn new tasks, while stable enough not to forget acquired knowledge of previous tasks; in a severe case, the latter situation is called catastrophic forgetting. Replay-based CL methods try to retain the performance by rehearsing stored raw input data or generated samples of the previous tasks. Another challenging situation for ML agents is to gain an acceptable overall performance on skewed data distribution of imbalanced datasets. In this thesis, for the first time in the literature, we analyze the performance of replay-based CL methods on imbalanced datasets in class-incremental scenario. Moreover, we suggest a new application of an adaptive sample re-weighting strategy called Meta-Weight Net for replay-based CL methods. Meta-Weight Net trains a NN to estimate the sample weights based on their respective loss value, which will define the adaptive weighted loss function. Our experiments show that using this strategy not only improves performance over an imbalanced dataset but also can help with more complex datasets where the data is balanced.