Abstract:

This study attempts to improve the Dark Experience Replay (DER) algorithm as a method to combat forgetting in continuous learning in neural networks. Although DER performs well in many standard continuous learning metrics, it may have limited effectiveness due to the use of pooled sampling and its limitations in certain scenarios with severe data changes and dynamic environments. Also, if the replay buffer is not large or diverse enough, there is a risk of catastrophic forgetting. In order to overcome these limitations, a new approach based on stratified random sampling is presented. This approach maintains the diversity and breadth of samples in the replay buffer and reduces forgetting. In order to evaluate the performance of the proposed method, the standard Split MNIST dataset was used.

The evaluation results show that the proposed method has higher accuracy than the DER algorithm with any buffer size. However, the speed of the DER algorithm is faster at larger replay buffers, while the proposed method is faster at smaller buffer sizes. Overall, the proposed method has lower error and higher accuracy than the DER algorithm. The best result with a replay buffer size of 100 shows that the proposed algorithm is 10.1% more accurate and 19.7% faster than DER.