Mitigating Forgetting in Continual Learning with Selective Gradient Projection
New algorithm reduces catastrophic forgetting in AI with 90% less memory overhead.
A research team led by Anika Singh, Aayush Dhaulakhandi, and Varun Chopade has introduced Selective Forgetting-Aware Optimization (SFAO), a breakthrough method addressing catastrophic forgetting in neural networks. This persistent challenge occurs when AI models overwrite previously learned information while adapting to new tasks, severely degrading performance on earlier ones. SFAO dynamically regulates gradient updates using cosine similarity analysis and a per-layer gating mechanism, allowing it to selectively project, accept, or discard parameter changes. This tunable approach employs efficient Monte Carlo approximation to balance plasticity (learning new tasks) with stability (retaining old knowledge), offering a more controlled alternative to traditional methods.
Experiments on standard continual learning benchmarks demonstrate SFAO's practical advantages. The method maintains competitive task accuracy while achieving a dramatic 90% reduction in memory overhead compared to existing techniques. This efficiency gain is particularly significant for deployment in resource-constrained scenarios like edge devices and mobile applications. By making continual learning more memory-efficient without sacrificing performance, SFAO paves the way for AI systems that can adapt sequentially to new data in real-world, dynamic environments—from personalized assistants to autonomous systems that evolve over time.
- SFAO reduces memory cost by 90% compared to previous continual learning methods
- Uses selective gradient projection with cosine similarity and per-layer gating to control forgetting
- Achieves competitive accuracy on benchmarks like MNIST while balancing plasticity and stability
Why It Matters
Enables efficient AI that learns continuously on devices with limited memory, from phones to sensors.