Research map / Learning theory and algorithms / Matrix and tensor methods
Multitask Spectral Learning of Weighted Automata
Guillaume Rabusseau, Borja Balle, Joelle Pineau
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 multitask learning (MT-SL) with all other languages as related tasks | SL · UNIDEP train size 100 | perplexity rel. improvement: 7.0744 | Improved |
| Restrict related tasks to the 4 closest languages | MT-SL · UNIDEP train size 1000 | perplexity rel. improvement: 3.2879 -> 4.4049 | Improved |
| Restrict related tasks to the 4 closest languages | MT-SL · UNIDEP train size 5000 | perplexity rel. improvement: 3.4187 -> 2.9689 | No clear effect |
| Restrict related tasks to the 4 closest languages | MT-SL · UNIDEP train size 100 | WER rel. improvement: 1.4919 -> 2.0883 | Improved |
| Add projection step (ablation) | MT-SL-noproj · UNIDEP all data, all languages | WER rel. improvement: −9.600 -> 1.4932 | Improved |
| Add projection step (ablation) | MT-SL-noproj · UNIDEP all data, all languages | perplexity rel. improvement: 0.6958 -> 3.1574 | Improved |
| Bag all languages' data into one SL training set instead of MT-SL | SL · UNIDEP train size 100, all languages | perplexity rel. improvement: −19.00 | Hurt |
| Bag data of 4 closest languages | SL · UNIDEP all data, 4 closest | WER rel. improvement: −3.588 | Hurt |
| Use MT-SL with 2, 4 or 8 related tasks instead of 2 | MT-SL · synthetic dS=10, dT=0 and 5 | perplexity and WER | Improved |
| Use MT-SL when tasks are loosely related (dS = dT = 10) | SL · synthetic | perplexity and WER | No clear effect |
| Use MT-SL on small task-specific dimension (dT = 0, 5) | SL · synthetic | perplexity and WER | Improved |
| Use MT-SL with 4 closest languages | SL · Basque, train size 500 | WER: ∼76% -> ∼70% -> ∼65% | Improved |
Closest papers by meaning
- Spectral Learning from a Single Trajectory under Finite-State Policies (ICML 2017 · similarity 0.87)
- Automata Cascades: Expressivity and Sample Complexity (AAAI 2023 · similarity 0.85)
- Weighted Automata Extraction from Recurrent Neural Networks via Regression on State Spaces (AAAI 2020 · similarity 0.84)
- Task Generalization with Autoregressive Compositional Structure: Can Learning from $D$ Tasks Generalize to $D^T$ Tasks? (ICML 2025 · similarity 0.82)
- Compositional Automata Embeddings for Goal-Conditioned Reinforcement Learning (NeurIPS 2024 · similarity 0.82)
- Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · similarity 0.82)
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.
