English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Class-incremental learning (CIL) is a challenging task that involvescontinually learning to categorize classes into new tasks without forgettingpreviously learned information. The advent of the large pre-trained models(PTMs) has fast-tracked the progress in CIL due to the highly transferable PTMrepresentations, where tuning a small set of parameters results instate-of-the-art performance when compared with the traditional CIL methodsthat are trained from scratch. However, repeated fine-tuning on each taskdestroys the rich representations of the PTMs and further leads to forgettingprevious tasks. To strike a balance between the stability and plasticity ofPTMs for CIL, we propose a novel perspective of eliminating training on everynew task and instead performing test-time adaptation (TTA) directly on the testinstances. Concretely, we propose "Test-Time Adaptation for Class-IncrementalLearning" (TTACIL) that first fine-tunes Layer Norm parameters of the PTM oneach test instance for learning task-specific features, and then resets themback to the base model to preserve stability. As a consequence, TTACIL does notundergo any forgetting, while benefiting each task with the rich PTM features.Additionally, by design, our method is robust to common data corruptions. OurTTACIL outperforms several state-of-the-art CIL methods when evaluated onmultiple CIL benchmarks under both clean and corrupted data.

Rethinking Class-incremental Learning in the Era of Large Pre-trained Models via Test-Time Adaptation