Reproduce “Dual-Model Distillation for Efficient Action Classification with Hybrid Edge-Cloud Solution” 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
Timothy Wei, Hsien Xin Peng, Elaine Xu, Bryan Zhao, Lei Ding, Diji Yang
A hybrid edge-cloud solution employing Dual-Model Distillation (DMD) is presented for efficient action classification, which achieved an F1 score of 92.1%—surpassing the large cloud model alone by 4.6%—while concurrently reducing energy consumption by 39.5% and runtime by 38.9%.
Key insights
- Intelligent orchestration via a switcher model can allow a hybrid edge-cloud system to achieve superior performance compared to even a standalone large cloud model, challenging the traditional accuracy-efficiency trade-off.
- Dual-Model Distillation (DMD) provides an effective unsupervised method for generating training data for a decision-making 'switcher' model, leveraging the agreement and disagreement between a small edge model and a large cloud model.
- Computational efficiency (reduced energy and runtime) and high accuracy can be simultaneously achieved by strategically offloading only when the local model is likely to diverge from the more powerful cloud model's prediction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.