Research map / Learning theory and algorithms / Matrix and tensor methods
Reverse-mode differentiation in arbitrary tensor network format: with application to supervised learning
Alex A. Gorodetsky, Cosmin Safta, John D. Jakeman
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 gradient-based nonlinear least squares instead of ALS to fit the tensor network | tensor network recovery (5x5 matrix, rank 2) · two-dimensional function recovery | probability of recovery: recovery using NLS occurs starting at 15 samples while ALS is delayed until 20 samples | Improved |
| use NLS instead of ALS on an over-parameterized 3x3 matrix, rank 3 | tensor network recovery · two-dimensional function recovery | probability of recovery: both approaches recover the function once nine data points are obtained | No clear effect |
| use NLS instead of ALS on an over-parameterized 5x5 matrix, rank 4 | tensor network recovery · two-dimensional function recovery | probability of recovery: NLS immediately fits well once the 25 sample threshold is passed | Improved |
| use NLS instead of ALS on balanced tree network (pi=2, r=2) | tree-tensor network · four-dimensional function recovery | probability of recovery: both ALS and NLS converge almost immediately after observing 16 data points | No clear effect |
| use NLS instead of ALS on balanced tree network (pi=3, r=3) | tree-tensor network · four-dimensional function recovery | samples needed for recovery: NLS does seem to converge 10 samples earlier than the ALS | Improved |
| use NLS instead of ALS on balanced tree network (pi=4, r=2, 52 parameters) | tree-tensor network · four-dimensional function recovery | probability of recovery: earlier transition into convergence by NLS, around approximately 50 samples | Improved |
| use NLS instead of ALS on non-traditional 5-D network (410 unknowns) | non-traditional tensor network · five-dimensional function recovery | samples needed for recovery: begins to converge almost one hundred samples earlier than the ALS solver | Improved |
| use NLS instead of ALS on 16-D unbalanced tree network (745 parameters) | non-traditional tensor network · sixteen-dimensional function recovery | samples needed for recovery: ALS needs several more times the number of samples to converge; ALS only converges to a level of 10-5 | Improved |
| use hierarchical Tucker with gradient-based fitting instead of functional tensor train | regression on real datasets · Housing, Speech, Music, Airfoil, Forestfires | validation MSE: Music 0.72094 -> 0.6536; Airfoil 0.46920 -> 0.40708; Forestfires 0.39465 -> 0.352 | Improved |
| use hierarchical Tucker with gradient-based fitting instead of functional tensor train | regression on real datasets · Skillcraft, Telemonit, Propulsion | validation MSE: Telemonit 0.04040 -> 0.079516; Propulsion 0.00009 -> 0.0006952; Skillcraft 0.54434 -> 0.70944 | Hurt |
| use hierarchical Tucker with gradient-based fitting instead of functional tensor train | regression on real datasets · CCPP (d=59, n=2000) | validation MSE: 0.06631 -> 0.06843 | No clear effect |
| use hierarchical Tucker with gradient-based fitting instead of functional tensor train | regression on real datasets · Galaxy (d=20, n=2000) | validation MSE: 0.00056 -> 0.000017 | Improved |
Closest papers by meaning
- Tensor Regression Networks (JMLR 2020 · similarity 0.88)
- Convolutions and More as Einsum: A Tensor Network Perspective with Advances for Second-Order Methods (NeurIPS 2024 · similarity 0.88)
- tnGPS: Discovering Unknown Tensor Network Structure Search Algorithms via Large Language Models (LLMs) (ICML 2024 · similarity 0.86)
- Low-Rank Tensor Transitions (LoRT) for Transferable Tensor Regression (ICML 2025 · similarity 0.86)
- Deep Tree Tensor Networks (NeurIPS 2025 · similarity 0.86)
- Geometry-aware training of factorized layers in tensor Tucker format (NeurIPS 2024 · 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.
