Long-Tail Zero and Few-Shot Learning via Contrastive Pretraining on and for Small Data
Publikation: Bidrag til tidsskrift › Tidsskriftartikel › Forskning › fagfællebedømt
Dokumenter
- Fulltext
Forlagets udgivne version, 1,13 MB, PDF-dokument
Preserving long-tail, minority information during model compression has been linked to algorithmic fairness considerations. However, this assumes that large models capture long-tail information and smaller ones do not, which raises two questions. One, how well do large pretrained language models encode long-tail information? Two, how can small language models be made to better capture long-tail information, without requiring a compression step? First, we study the performance of pretrained Transformers on a challenging new long-tail, web text classification task. Second, to train small long-tail capture models we propose a contrastive training objective that unifies self-supervised pretraining, and supervised long-tail fine-tuning, which markedly increases tail data-efficiency and tail prediction performance. Third, we analyze the resulting long-tail learning capabilities under zero-shot, few-shot and full supervision conditions, and study the performance impact of model size and self-supervision signal amount. We find that large pretrained language models do not guarantee long-tail retention and that much smaller, contrastively pretrained models better retain long-tail information while gaining data and compute efficiency. This demonstrates that model compression may not be the go-to method for obtaining good long-tail performance from compact models.
Originalsprog | Engelsk |
---|---|
Artikelnummer | 10 |
Tidsskrift | Computer Sciences & Mathematics Forum |
Vol/bind | 3 |
Antal sider | 18 |
ISSN | 2813-0324 |
DOI | |
Status | Udgivet - 2022 |
ID: 339336056