Overcoming Catastrophic Forgetting: From Efficiency to Safety

Overview

The increasing use of artificial intelligence (AI) in education, healthcare, and infrastructure requires systems that can adapt and learn continuously. A central obstacle is catastrophic forgetting, where models lose previously learned knowledge when trained on new tasks. This thesis develops and evaluates continual learning methods for two practical settings: resource-constrained scenarios and safety-critical applications where preserving alignment is essential. We make three main contributions. First, we introduce a lightweight regularizer that improves retention under rehearsal with minimal cost. Second, we present SMILE; a video continual-learning method that greatly reduces memory use while improving performance. Third, we adapt continual learning (CL) approaches to preserve safety alignment as large language models are fine-tuned and show that CL methods can prevent alignment degradation and match or exceed existing safety approaches. Collectively, these contributions lead to effective solutions for both efficient adaptation and safety preservation in deployed AI systems.