Research map / Learning theory and algorithms / Matrix and tensor methods
DeepTensor: Low-Rank Tensor Decomposition With Deep Network Priors
Vishwanath Saragadam, Randall Balestriero, Ashok Veeraraghavan, Richard G. Baraniuk
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 fixed schedule instead of cosine annealing at max LR 1e-2 (ablation) | DeepTensor · toy matrix low-rank approximation | PSNR (dB): 16.5 -> 17.6 | Improved |
| use max learning rate 1e-4 or higher instead of 1e-5 (fixed schedule) (ablation) | DeepTensor · toy matrix low-rank approximation | PSNR (dB): 16.5 -> 17.6 | Improved |
| replace SVD with DeepTensor (ablation) | SVD · toy matrix low-rank approximation | PSNR (dB): 14.4 -> 17.6 | Improved |
| use ReLU as final activation instead of softplus (ablation) | DeepTensor NMF · MNIST NMF | PSNR (dB): 7.4 -> 7.6 | Improved |
| use ReLU as final activation instead of softplus (ablation) | DeepTensor NMF · CIFAR10 NMF | PSNR (dB): 8.4 -> 8.8 | Improved |
| replace classical NMF with DeepTensor (ReLU) (ablation) | NMF · MNIST NMF | PSNR (dB): 7.5 -> 7.6 | Improved |
| replace classical NMF with DeepTensor (ReLU) | NMF · CIFAR10 NMF | PSNR (dB): 8.2 -> 8.8 | Improved |
| replace baseline NMF with DeepTensor | NMF · GoogleCommands spectrograms NMF | PSNR (dB): 4.6 -> 9.0 | Improved |
| replace baseline semi-NMF with DeepTensor | NMF · GoogleCommands spectrograms semi-NMF | PSNR (dB): 4.7 -> 8.8 | Improved |
| replace SVD with DeepTensor under Poisson noise | SVD · hyperspectral denoising (Poisson, rank 20) | PSNR (dB): approximately 6dB increase | Improved |
| replace BM3D with DeepTensor under Poisson noise | BM3D · hyperspectral denoising (Poisson, rank 20) | PSNR (dB): 3dB over BM3D | Improved |
| use DeepTensor instead of single 3D-conv DN for tensor decomposition | single DN with 3D convolutions · 3D tensor decomposition | training time: 60 times faster | Improved |
Closest papers by meaning
- Attention-Guided Low-Rank Tensor Completion (TPAMI 2024 · similarity 0.89)
- Deep Rank-One Tensor Functional Factorization for Multi-Dimensional Data Recovery (AAAI 2025 · similarity 0.89)
- HLRTF: Hierarchical Low-Rank Tensor Factorization for Inverse Problems in Multi-Dimensional Imaging (CVPR 2022 · similarity 0.88)
- Stable Low-rank Tensor Decomposition for Compression of Convolutional Neural Network (ECCV 2020 · similarity 0.88)
- Distribution-Aware Tensor Decomposition for Compression of Convolutional Neural Networks (NeurIPS 2025 · similarity 0.87)
- Low-Rank Tensor Transitions (LoRT) for Transferable Tensor Regression (ICML 2025 · similarity 0.86)
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.
