Reproduce “Re-Evaluating Continual Learning with Few-Shot Adaptation” 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
Amogh Inamdar, Matthew So, Vici Milenia, Richard Zemel
The authors evaluate few-shot recovery in continual image classification and propose foresight meta-learning, which adjusts a model’s starting point using examples from upcoming tasks. On backward tasks, five-shot adaptation often raises SGD accuracy substantially from near-random zero-shot levels, approaching replay methods. Foresight improves forward adaptation in most settings, but results depend on task sequence and require future-task examples.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.