Reproduce “SelectiveNet: A Deep Neural Network with an Integrated Reject Option” 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
Yonatan Geifman, Ran El-Yaniv
SelectiveNet introduces a deep neural network architecture that integrates prediction and rejection functionalities, optimizing both simultaneously for a user-defined coverage constraint. This approach consistently yields improved risk-coverage trade-offs, showing selective risk reductions up to 48.16% in classification and 17.09% in regression over baseline methods while offering computational efficiency for selective regression.
Key insights
- End-to-end optimization of both prediction and selection functions within a single network leads to superior risk-coverage trade-offs compared to post-hoc rejection strategies.
- Explicitly training for selective prediction enables the network to allocate representational capacity more efficiently, forming highly discriminative features for covered instances while allowing rejected instances to be less separated.
- The auxiliary prediction head plays a critical role during training by ensuring the main body block processes all instances, promoting the learning of robust, general-purpose features before selective aspects dominate.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.