Incremental Prototype Tuning for Class Incremental Learning

Abstract

Class incremental learning has attracted much attention, but most existing works still continually fine-tune the entire representation model, inevitably resulting in much catastrophic forgetting. Instead of struggling to fight against such forgetting by replaying or distillation like most of the existing methods, we take a novel pre-train-and-prompt-tuning paradigm to sequentially learn new visual concepts based on a fixed semantic-rich pre-trained representation model. In detail, we incrementally prompt-tune category prototypes for classification and example prototypes to compensate for semantic drift, the problem caused by learning bias at different phases. Extensive experiments conducted on the mainstream incremental learning benchmarks demonstrate that our method outperforms other state-of-theart methods.

S-Prompts Learning with Pre-trained Transformers: An Occam's Razor for Domain Incremental Learning

“…There have been also emerging some other transformer-based incremental learning methods like (71; 27; 37), and other incremental prompting methods like (42; 65) 12 methods without prompting demand much more tuning on the whole network than the prompting methods like (76) and ours. Also, they keep applying traditional dependent learning paradigm to play the tug-of-war game.…”