Overcoming Network Saturation in Continual Learning: A Method for Dynamic Parameter Adjustment

More Info
expand_more

Abstract

Neural networks have made significant progress in domains like image recognition and natural language processing. However, they encounter the challenge of catastrophic forgetting in continual learning tasks, where they sequentially learn from distinct datasets. Learning a new task can lead to forgetting important information from previous tasks, resulting in decreased performance on those earlier tasks. This issue is further intensified in dynamic scenarios where the task sequence varies unpredictably. To address this problem, architectural methods have been developed to modify a neural network's structure, creating or adapting subnetworks to retain task-specific knowledge and mitigate catastrophic forgetting. However, these solutions can lead to network saturation, where the accumulation of task-specific adaptations hampers the network's ability to learn new tasks. This research aims to address the problem of network saturation by developing innovative methods that enable neural networks to maintain high performance across both existing and new tasks in continual learning scenarios. Eventually, the new model improved its learning ability on new tasks in the presence of an allowable forgetting, while demonstrating better overall learning ability.

Files

Xusen_Master_Thesis.pdf
(pdf | 9.9 Mb)
Unknown license