Research map / Label-efficient and robust learning / Recognition and long-tail
Auxiliary Learning by Implicit Differentiation
Aviv Navon, Idan Achituve, Haggai Maron, Gal Chechik, Ethan Fetaya
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 |
|---|---|---|---|
| Learn nonlinear combinations of predefined auxiliary losses (ablation) | AuxiLearn on CUB-200-2011 · CUB 200-2011 5-shot | Top 1 accuracy: 41.71 ± 0.4 -> 47.07 ± 0.1 | Improved |
| Use a deep linear network instead of a shallow linear loss combiner (ablation) | AuxiLearn on CUB-200-2011 · CUB 200-2011 5-shot | Top 1 accuracy: 41.71 ± 0.4 -> 45.84 ± 0.3 | Improved |
| Apply a CNN to spatial maps of per-pixel task losses (ablation) | AuxiLearn on NYUv2 · NYUv2 | mIoU: 20.04 ± 0.38 -> 20.54 ± 0.30 | Improved |
| Learn auxiliary labels using an auxiliary set to optimize for main-task generalization | AuxiLearn · CIFAR10 5% | Accuracy: 58.2 ± 0.3 -> 60.7 ± 1.3 | Improved |
| Constrain the loss-combination network to be monotone (ablation) | AuxiLearn on CUB-200-2011 · CUB 200-2011 5-shot | Top 1 accuracy: 46.3 ± 0.32 -> 47.07 ± 0.10 | Improved |
| Constrain the loss-combination network to be monotone (ablation) | AuxiLearn on CUB-200-2011 · CUB 200-2011 10-shot | Top 1 accuracy: 58.84 ± 0.04 -> 59.04 ± 0.22 | Improved |
| Tune auxiliary parameters on the training set rather than a separate auxiliary set (ablation) | AuxiLearn · Illustrative regression with helpful and harmful auxiliaries | Generalization | Hurt |
| Keep the learned auxiliary network fixed during primary-network retraining instead of jointly updating it (ablation) | AuxiLearn on CUB-200-2011 · CUB 200-2011 5-shot | Accuracy: 46.7% compared to 40% | Hurt |
| Add an auxiliary-label teacher learned with AuxiLearn rather than using MAXL | AuxiLearn on Oxford-IIIT Pet · Oxford-IIIT Pet 30-shot | Accuracy: 26.3 ± 0.6 -> 37.0 ± 0.6 | Improved |
| Increase the depth of the auxiliary network (ablation) | AuxiLearn on CUB-200-2011 · CUB 200-2011 5-shot | Accuracy | Improved |
| Allocate too many labeled examples from the training set to the auxiliary set (ablation) | AuxiLearn on CUB-200-2011 · CUB 200-2011 5-shot | Accuracy | Hurt |
| Use a nonlinear loss combiner instead of a linear one (ablation) | AuxiLearn on NYUv2 · NYUv2 | mIoU: 20.04 ± 0.38 -> 20.09 ± 0.34 | No clear effect |
Closest papers by meaning
- AUXILIARY TASK UPDATE DECOMPOSITION: THE GOOD, THE BAD AND THE NEUTRAL (ICLR 2021 · similarity 0.87)
- Aux-NAS: Exploiting Auxiliary Labels with Negligibly Extra Inference Cost (ICLR 2024 · similarity 0.87)
- Self-Supervised Generalisation with Meta Auxiliary Learning (NeurIPS 2019 · similarity 0.87)
- Joint Data-Task Generation for Auxiliary Learning (NeurIPS 2023 · similarity 0.86)
- Auxiliary Learning with Joint Task and Data Scheduling (ICML 2022 · similarity 0.85)
- Inductive biases of multi-task learning and finetuning: multiple regimes of feature reuse (NeurIPS 2024 · similarity 0.83)
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.
