Research map / Architectures and efficiency / RNNs and normalization
HALO: Hardware-Aware Learning to Optimize
Chaojian Li, Tianlong Chen, Haoran You, Zhangyang Wang, Yingyan Lin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Jacobian regularizer to vanilla HALO (ablation) | HALO · CIFAR-10-A wider optimizee | adaptation accuracy after convergence | Improved |
| add structural sparsity regularizer to HALO with Jacobian regularizer (ablation) | HALO · CIFAR-10-A wider optimizee | accuracy under same energy: up to 11.70% higher accuracy | Improved |
| add structural sparsity regularizer to HALO with Jacobian regularizer (ablation) | HALO · CIFAR-10-A wider optimizee | training loss under same energy: up to 16.72% lower | Improved |
| add both regularizers to vanilla HALO to beat Adam (ablation) | HALO · CIFAR-10-A wider optimizee | loss and accuracy after convergence | Improved |
| add structural sparsity regularizer to Adam (ablation) | Adam · CIFAR-10-A wider optimizee | adaptation accuracy | Hurt |
| use progressively increased update probability 10%-30%-50% instead of 50%-30%-10% (ablation) | HALO · CIFAR-10-A wider optimizee | accuracy under same energy: up to 20.28% Acc | Improved |
| use progressively increased update probability instead of uniform 30%-30%-30% (ablation) | HALO · CIFAR-10-A wider optimizee | accuracy under same energy | Improved |
| use HALO instead of Hierarchical-L2O | HALO · TDP | final test accuracy / energy: 80.28 vs 79.19; 31.8% less energy | Improved |
| use HALO instead of Hierarchical-L2O | HALO · GSAD | final test accuracy / energy: 87.00 vs 80.31; 58.6% - 60.0% less energy | Improved |
| use HALO instead of Hierarchical-L2O | HALO · SP | final test accuracy / energy: 98.16 vs 97.20; 44.3% - 45.1% less energy | Improved |
| use HALO instead of Adam | HALO · TDP | energy at 3k iterations: 0.13 -> 0.45 kJ | Hurt |
Closest papers by meaning
- Boosted Trees on a Diet: Compact Models for Resource-Constrained Devices (ICLR 2026 · similarity 0.86)
- Constrained deep neural network architecture search for IoT devices accounting for hardware calibration (NeurIPS 2019 · similarity 0.86)
- TinyTrain: Resource-Aware Task-Adaptive Sparse Training of DNNs at the Data-Scarce Edge (ICML 2024 · similarity 0.86)
- On-Device Training Under 256KB Memory (NeurIPS 2022 · similarity 0.85)
- TinyTTA: Efficient Test-time Adaptation via Early-exit Ensembles on Edge Devices (NeurIPS 2024 · similarity 0.84)
- The ML.ENERGY Benchmark: Toward Automated Inference Energy Measurement and Optimization (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
