Reproduce “PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruning” 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
Arun Mallya, Svetlana Lazebnik
Researchers at the University of Illinois Urbana-Champaign developed PackNet, a method that enables a single deep neural network to sequentially learn multiple tasks by iteratively pruning and re-allocating network parameters. This approach maintains performance on prior tasks, achieving accuracy comparable to individually trained models (e.g., within 1-2% top-1 error), while significantly reducing storage overhead (e.g., 595 MB for four tasks versus 2,173 MB for individual networks).
Key insights
- Deep neural networks possess significant parameter redundancy that can be strategically leveraged to accommodate new tasks within a fixed model capacity.
- Iterative pruning and re-training allow for efficient re-allocation of network parameters, enabling the learning of diverse new tasks without increasing model size.
- By freezing parameters dedicated to learned tasks, PackNet provides strong robustness against catastrophic forgetting, maintaining performance on older tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.