Research map / Large language models / Language, safety and interpretability
AANG : Automating Auxiliary Learning
Lucio M. Dery, Paul Michel, Mikhail Khodak, Graham Neubig, Ameet Talwalkar
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 adaptive objective weighting rather than static weights (ablation) | AANG-TD · Five NLP tasks, average | reported average performance: 81.07 -> 82.21 | Improved |
| Use 24 generated objectives rather than combining the 3 named objectives | AANG · Five NLP tasks, average | reported average performance: 81.55 -> 82.21 | Improved |
| Add external in-domain data and expand the search space from AANG-TD to AANG-TD+ED | AANG-TD · ACL-ARC | F1: 73.26 -> 73.70 | No clear effect |
| Increase sampled objectives per iteration from 3 of 24 to 6 of 24 (ablation) | AANG-TD · ACL-ARC | F1: 72.11 -> 73.26 | Improved |
| Increase sampled objectives per iteration from 3 of 24 to 6 of 24 (ablation) | AANG-TD · SCIERC | F1: 82.35 -> 82.98 | Improved |
| Increase sampled objectives per iteration from 3 of 24 to 6 of 24 (ablation) | AANG-TD · SE-2016-6 | accuracy: 72.46 -> 72.46 | No clear effect |
| Increase sampled objectives per iteration from 3 of 24 to 6 of 24 (ablation) | AANG-TD · CHEMPROT | accuracy: 83.91 -> 83.69 | Hurt |
| Increase sampled objectives per iteration from 3 of 24 to 6 of 24 (ablation) | AANG-TD · H.PARTISAN | accuracy: 98.46 -> 97.95 | Hurt |
| Use end-task-aware multitasking instead of end-task-agnostic auxiliary training | Auxiliary learning on task data · Five NLP tasks, average | reported average performance: 78.72 -> 79.84 | Improved |
Closest papers by meaning
- Should We Be Pre-training? An Argument for End-task Aware Training as an Alternative (ICLR 2022 · similarity 0.90)
- Joint Data-Task Generation for Auxiliary Learning (NeurIPS 2023 · similarity 0.86)
- AUXILIARY TASK UPDATE DECOMPOSITION: THE GOOD, THE BAD AND THE NEUTRAL (ICLR 2021 · similarity 0.86)
- Aux-NAS: Exploiting Auxiliary Labels with Negligibly Extra Inference Cost (ICLR 2024 · similarity 0.86)
- Prefer to Classify: Improving Text Classifiers via Auxiliary Preference Learning (ICML 2023 · similarity 0.85)
- A Mathematical Exploration of Why Language Models Help Solve Downstream Tasks (ICLR 2021 · similarity 0.85)
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.
