Reproduce “Factorizing Knowledge in Neural Networks” 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
Xingyi Yang, Jingwen Ye, Xinchao Wang
Knowledge Factorization (KF) divides a pretrained teacher’s knowledge into a shared representation and task-specific components that users can combine without retraining the assembled model. The authors train these factors to imitate the teacher and meet task objectives, using an InfoMax Bottleneck to encourage separation. On ImageNet’s 11-task setup, KF reaches 72.20% top-1 accuracy; no matched 11-task KD result is reported.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.