LG AI Research 63

Most deep learning models learn by processing large volumes of data in parallel, but in security or low-power environments, situations frequently arise where data must be learned in a divided manner, known as Continual Learning. In this case, learning new data causes a phenomenon called Catastrophic Forgetting, where previous information is forgotten, and resolving this is the core goal of continual learning.

LG AI Research and Professor Scott Sanner's team at the University of Toronto proposed the ASER (Adversarial Shapley Value Experience Replay) algorithm, which minimizes forgetting in the Online Class-incremental Setting environment. This method is the first to apply KNN Shapley value, a metric that evaluates the importance of data, to the continual learning problem.

The key features of ASER are as follows:

• Utilization of KNN Shapley value: Evaluates where each data point is located within the Feature Space and how much it influences the classification of other data.
• Strategic memory update: Prioritizes keeping data with high positive values in memory, helping stable learning for each Class.
• Efficient memory retrieval: Prevents distortion of existing knowledge by selecting data of other classes that lie at the boundary with new data, or by selecting data that well represents the corresponding class for learning.

Experimental results showed that ASER achieved higher accuracy than existing methods such as Random Replay or MIR, and demonstrated the performance of continual learning by more effectively utilizing the distribution of data.