This paper introduces Visual Contrastive Self-Distillation, a method that removes the need for external teacher information and privileged answers in on-policy self-distillation, allowing for simpler and more efficient learning. Practitioners might care about this approach because it can lead to better performance in language models.
Firehose
Filtered to Papers, tagged “input-based learning” · clear filters
Browse: People · Companies · Papers · Podcasts · Hacker News