Research map / Architectures and efficiency / RNNs and normalization
Constrained Optimization Based Low-Rank Approximation of Deep Neural Networks
Chong Li, C. J. Richard Shi
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 |
|---|---|---|---|
| use COBLA instead of POS-CVX after fine-tuning (budget 0.7, Equ. 3) | SqueezeNet low-rank · ImageNet val | top-1 accuracy (%): 55.7% vs baseline -2.0% | Improved |
| use COBLA instead of POS-CVX (budget 0.5, Equ. 3) | SqueezeNet low-rank · ImageNet val | top-1 accuracy (%): 52.6% vs baseline -7.4% | Improved |
| use COBLA instead of POS-CVX (budget 0.5, Equ. 4) | SqueezeNet low-rank · ImageNet val | top-5 accuracy (%): 76.2% vs baseline -4.1% | Improved |
| use COBLA instead of POS-CVX (budget 0.6, Equ. 4) | SqueezeNet low-rank · ImageNet val | top-1 accuracy (%): 54.3% vs baseline -3.8% | Improved |
| use COBLA instead of R-VBMF under severe budget, 10 epoch fine-tuning | SqueezeNet low-rank · ImageNet val | top-1 accuracy (%): 15.9% vs near chance | Improved |
| use Equ. 3 decomposition instead of Equ. 4 (before fine-tuning) (ablation) | COBLA · ImageNet val | top-1 accuracy (%): COBLA much better with Equ. 3 before fine-tuning | Improved |
| switch decomposition scheme from Equ. 4 to Equ. 3 after fine-tuning (ablation) | COBLA · ImageNet val | top-1 accuracy (%): difference within 1% | No clear effect |
| use sub-sampled gradient (5% of data) instead of exact gradient in SQP (ablation) | COBLA · CIFAR-10 NIN | top-1 accuracy (%): within 1% | No clear effect |
| use COBLA allocation instead of manual trial-and-error ranks [27] | VGG-16 low-rank · ImageNet val | top-5 accuracy (%): 89.8% -> 89.8% with 12% less cost | Improved |
| push COBLA to 30% cost reduction versus [27] | VGG-16 low-rank · ImageNet val | top-5 accuracy (%): 89.8% -> 88.9% | Hurt |
| add data-dependent decomposition (ablation) | low-rank approximation · ImageNet val | accuracy | No clear effect |
Closest papers by meaning
- COALA: Numerically Stable and Efficient Framework for Context-Aware Low-Rank Approximation (NeurIPS 2025 · similarity 0.86)
- Approximation to Smooth Functions by Low-Rank Swish Networks (ICML 2025 · similarity 0.86)
- Robust low-rank training via approximate orthonormal constraints (NeurIPS 2023 · similarity 0.85)
- Explicitly Imposing Constraints in Deep Networks via Conditional Gradients Gives Improved Generalization and Faster Convergence (AAAI 2019 · similarity 0.85)
- Compressing Neural Networks: Towards Determining the Optimal Layer-wise Decomposition (NeurIPS 2021 · similarity 0.85)
- Distribution-Aware Tensor Decomposition for Compression of Convolutional Neural Networks (NeurIPS 2025 · similarity 0.84)
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.
