Reproduce “BEEM: Boosting Performance of Early Exit DNNs using Multi-Exit Classifiers as Experts” 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
Divya Jyoti Bajpai, Manjesh Kumar Hanawal
The BEEM framework enhances early exit Deep Neural Networks by aggregating weighted confidence from intermediate classifiers and dynamically accounting for prediction consistency. This method delivers a 1.5x to 2.1x speed-up, reaching up to 3.33x for large models, while maintaining or improving accuracy over the full network, particularly on simpler tasks.
Key insights
- Aggregating weighted confidence from intermediate classifiers, treating them as "experts," leads to more robust early exit decisions.
- Incorporating dynamic checks for prediction consistency prevents premature exits due to transient confidence, enhancing overall reliability.
- An optimization-based thresholding strategy can ensure early exits perform at least as well as the full network, effectively mitigating the accuracy-speed trade-off.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.