Reproduce “Rethinking Experience Replay: a Bag of Tricks for Continual Learning” on OpenResearch

About this paper

Pietro Buzzega, Matteo Boschini, Angelo Porrello, Simone Calderara

This work enhances basic Experience Replay (ER) with a "bag of tricks," creating ER+T, a robust and efficient approach that addresses key limitations of naive ER in Class-Incremental Learning. ER+T consistently surpasses the performance of more complex state-of-the-art rehearsal-based methods across diverse continual learning benchmarks, demonstrating significant accuracy improvements while maintaining competitive computational efficiency.

Key insights

• Simple, well-tuned modifications to basic Experience Replay (ER) can yield superior performance compared to more complex Continual Learning (CL) algorithms.
• Addressing specific limitations of naive ER, such as overfitting, prediction bias towards newer classes, and suboptimal buffer management, is crucial for improving its effectiveness.
• Some proposed tricks, like Bias Control and Exponential Learning Rate Decay, are generalizable and can also enhance non-rehearsal-based CL methods.

Summary by alphaXiv.