New AI method SFAO reduces memory usage by 90% while preventing neural networks from forgetting previously learned tasks

Researchers propose Selective Forgetting-Aware Optimization (SFAO) to address catastrophic forgetting, where neural networks overwrite old knowledge when learning new tasks

SFAO uses cosine similarity and per-layer gating to selectively project, accept, or discard gradient updates while balancing learning flexibility and stability

Tested on continual learning benchmarks including MNIST, SFAO achieves competitive accuracy with 90% reduction in memory requirements compared to existing methods

The method uses efficient Monte Carlo approximation to make the selective gating mechanism computationally practical for resource-constrained environments