Reproduce “CLIP model is an Efficient Continual Learner” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Vishal Thengane, Salman Khan, Munawar Hayat, Fahad Khan
Continual-CLIP tests whether a frozen pretrained vision-language model can serve as a class-incremental learning baseline without task-wise fine-tuning or a dedicated replay buffer. It compares each image with text descriptions of available classes and predicts the closest match. On 50-step CIFAR-100, it reports 76.49% average and 66.72% final accuracy, but at 10 steps its average is slightly below DER without pruning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.