Reproduce “Unwrapping The Black Box of Deep ReLU Networks: Interpretability, Diagnostics, and Simplification” 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
Agus Sudjianto, William Knauth, Rahul Singh, Zebin Yang, Aijun Zhang
A framework is developed for exactly interpreting, diagnosing, and simplifying deep ReLU networks by leveraging their intrinsic piecewise linear nature. This approach "unwraps" models into local linear models, revealing decision processes and enabling architectural simplification that can improve predictive performance, as seen with an AUC increase from 0.8316 to 0.8388 on a financial dataset.
Key insights
- Deep ReLU networks are intrinsically piecewise linear functions, where each distinct activation pattern defines a specific linear region in the input space.
- Over-parameterized networks often contain numerous 'single-instance' or 'single-class' regions, indicating potential overfitting or redundant model complexity.
- Simplifying deep networks by merging similar local linear models can lead to improved generalization performance and significantly reduced architectural complexity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.