Research map / Detection and segmentation / CNN features
A Two-Stage Learning-to-Defer Approach for Multi-Task Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a two-stage learning-to-defer framework for joint classification and regression with multiple experts, with a cross-entropy surrogate shown to be Bayes- and (G,R)-consistent, and evaluates it on Pascal VOC object detection and MIMIC-IV EHR tasks.
Problem
Learning-to-defer has been studied for classification or regression separately, but many real tasks (e.g., object detection) need joint classification and regression deferral to a selected agent in a two-stage (offline-trained model and experts) setting.
Approach
A cross-entropy-based surrogate deferral loss over J+1 agents with aggregated cost weights tau_j, with consistency bounds parameterized by nu and the L1 norm of expected costs, plus a closed-form minimizability gap for multiple experts and a multi-task cost that sums classification and regression losses.
Key results
- On the EHR task, the proposed rejector reaches 70.0% accuracy and 1.28 sL1, versus 71.3% / 1.45 for the classification-focused rejector and 50.7% / 1.18 for the regression-focused one.
- On the EHR task the proposed method allocates 66% of queries to the model, versus 60% and 38% for the two baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use joint classification and regression deferral instead of a classification-focused rejector | Two-stage L2D on MIMIC-IV · MIMIC-IV | sL1: 1.45 -> 1.28 | Helped |
| Use joint classification and regression deferral instead of a classification-focused rejector | Two-stage L2D on MIMIC-IV · MIMIC-IV | accuracy (%): 71.3 -> 70.0 | Hurt |
| Use joint classification and regression deferral instead of a regression-focused rejector | Two-stage L2D on MIMIC-IV · MIMIC-IV | accuracy (%): 50.7 -> 70.0 | Helped |
| Use joint classification and regression deferral instead of a regression-focused rejector | Two-stage L2D on MIMIC-IV · MIMIC-IV | sL1: 1.18 -> 1.28 | Hurt |
Surprising findings
- Regression-focused rejector gets the best sL1 but only 50.7% accuracy, showing single-task deferral trades off poorly across tasks.
Limitations
- Experiments are on only two tasks, with synthetic expert specialization assumed for EHR (experts correct on assigned clusters, uniform outputs elsewhere).
- The EHR experiment assumes beta_j=0 and simulated expert behavior rather than real experts.
Experimental setup
Datasets: Pascal VOC MIMIC-IV
Benchmarks: Pascal VOC MIMIC-IV
Metrics: mAP allocation ratio accuracy smooth L1 loss
Baselines: classification-focused two-stage L2D rejector regression-focused two-stage L2D rejector
Builds on: two-stage learning-to-defer (Mao et al.) score-based L2D surrogates H-consistency bounds comp-sum surrogate losses
Closest papers by meaning
- Exploiting Human-AI Dependence for Learning to Defer (ICML 2024 · similarity 0.87)
- Regression with Multi-Expert Deferral (ICML 2024 · similarity 0.87)
- Mastering Multiple-Expert Routing: Realizable $H$-Consistency and Strong Guarantees for Learning to Defer (ICML 2025 · similarity 0.85)
- Realizable $H$-Consistent and Bayes-Consistent Loss Functions for Learning to Defer (NeurIPS 2024 · similarity 0.85)
- Adversarial Robustness in Two-Stage Learning-to-Defer: Algorithms and Guarantees (ICML 2025 · similarity 0.85)
- A Unifying Post-Processing Framework for Multi-Objective Learn-to-Defer Problems (NeurIPS 2024 · 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.