LG AI Research: Research on Online Class-Incremental Continual Learning

Conventional image-based deep learning models use multiple GPUs to train on massive amounts of data, but in real-world environments they face constraints such as data security and limited computing resources. When data is learned sequentially, a phenomenon called Catastrophic Forgetting occurs, in which the model forgets existing knowledge upon exposure to new data.

To address this issue, the field of Continual Learning has emerged, and LG AI Research presented the ASER (Adversarial Shapley Value Experience Replay) algorithm at AAAI-21. This algorithm is the first to apply KNN-SV (K-Nearest Neighbor Shapley Value), a data valuation method, to continual learning, establishing a theoretical foundation for the approach.

Continual learning is broadly divided into two settings:

• Task-incremental setting: A relatively simple environment that utilizes both the current task's data and the task identifier (Task ID).
• Online-class incremental setting: A far more challenging and realistic environment in which the model learns in real time through a data stream without task information and must distinguish among all classes.

ASER is built on Experience Replay, the most promising methodology in this field. This approach stores some data from the data stream in memory and retrains it together with new data. By leveraging KNN-SV, ASER performs strategic and intuitive data selection during the memory update and retrieval process, maintaining higher accuracy than existing methods while minimizing forgetting.