Continual Learning for Classification Tasks

The traditional machine learning models can only learn from stationary data stream and are very vulnerable to the changing environment. To encourage the machine learners to be more human-like and practical in the real-world, continual learning was put forward and attracts a surge of attention in recent few years. Conventionally, there are three scenarios in continual learning which are class-incremental, task-incremental and domain-incremental and each scenario is associated with a specific model configuration. The most critical problem addressed by continual learning is catastrophic forgetting on previously learned tasks, whereupon this thesis aims to tackle this problem for classification tasks in different continual learning settings. Towards this end, this thesis first proposes a continual learning approach with dual regularizations on output and representation spaces for domain-incremental learning. The output space is regularized by means of knowledge distillation to preserve acquired information from old tasks and the representational regularization resorts to feature selection via sparse regularizer so that the unique classifier is able to perform well in the filtered feature space. The second approach is an online continual learning algorithm applicable to all three continual learning scenarios. Specifically, the feature propagation from previous feature space to current one is employed to retain some past knowledge and a contrastive regularization loss to prevent the feature space from drastic changes is further applied to strengthen such effect. For eliminating the domain shifts among different tasks, a supervised contrastive loss is leveraged as a complement to improve the separability of different classes. To consider the case of data insufficiency, this thesis further focuses on a more practical continual learning setting - few-shot class-incremental learning. Instead of optimizing the model on each task, we utilize a dictionary matrix to project the visual space to a new space for classification and only update the dictionary when learning new tasks. To further facilitate future adaptation, multiple pseudo classes are added into the initial task to equip the model with more compatibility and extensibility. All the proposed methods are extensively evaluated on a variety of benchmark datasets compared to other state-of-the-art methods, demonstrating the significant effectiveness and superiority.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.