Remembering by Reconstructing: Domain Incremental Learning With Test-Time Training on Video Streams
Researchers have developed a new method for domain incremental learning that embraces catastrophic forgetting rather than trying to prevent it. The approach uses domain-specific LoRA adapters and a self-supervised masked autoencoder head. At inference, test-time training on the MAE head helps identify the correct adapter to recall domain-specific knowledge, making it suitable for streaming data like video. AI
IMPACT This method could improve AI's ability to adapt to changing data streams in real-world applications like video analysis.