Continual Learning for Video Action Recognition using Parameter Allocation
Lukmanul Hakeem, C J Aromal, Sushant Kumar, Sumit Datta · 2024
Conventional machine learning approaches have shown great promise in video action recognition tasks. However, these approaches have a fatal flaw: where in they fail in dynamically adapting to new stimuli in an ever evolving environment without losing any of it's previously learned knowledge. This missing ability in general machine learning is called Continual learning, a key technique in attaining Artificial General Intelligence systems (AGI's) making them capable of adapting to real-world scenario's. Catastrophic forgetting is the obstacle that hinders machine learning models from achieving continual learning. There are various approaches for embedding continual learning ability in a standard neural network architecture. In computer vision, even though continual learning has been explored in tasks like image classification, object detection and video recognition, most studies explore a replay-based continual learning approach that has a computational overhead of having to store samples from previous tasks. Here, we propose a continual learning video action recognition architecture based on a Parameter Allocation approach. The proposed method can continually learn video recognition tasks without losing its previously acquired knowledge in comparison to the traditional models and is also free from storing exemplars from previous tasks.