Memory Consolidation Enables Long-Context Video Understanding
This work addresses the challenge of long-context video understanding for AI systems, offering an efficient solution that is incremental by building on existing pre-trained transformers.
The paper tackles the problem of limited temporal context in transformer-based video encoders by fine-tuning pre-trained models to attend to non-parametric memories from past activations, resulting in a new state-of-the-art on benchmarks like EgoSchema, Perception Test, and Diving48 while outperforming larger models.
Most transformer-based video encoders are limited to short temporal contexts due to their quadratic complexity. While various attempts have been made to extend this context, this has often come at the cost of both conceptual and computational complexity. We propose to instead re-purpose existing pre-trained video transformers by simply fine-tuning them to attend to memories derived non-parametrically from past activations. By leveraging redundancy reduction, our memory-consolidated vision transformer (MC-ViT) effortlessly extends its context far into the past and exhibits excellent scaling behavior when learning from longer videos. In doing so, MC-ViT sets a new state-of-the-art in long-context video understanding on EgoSchema, Perception Test, and Diving48, outperforming methods that benefit from orders of magnitude more parameters.