PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning to Defer with Limited Expert Predictions

Patrick Hemmer, Lukas Thede, Michael Vössing, Johannes Jakubik, Niklas Kühl

arXiv:2304.07306v1cs.LGcs.AIcs.HC

TL;DR

Learning to defer needs expert predictions as well as ground-truth labels, making it costly to train when experts change or predictions are expensive to acquire. The paper generates artificial expert predictions from a small set of expert-labeled examples. Across its evaluations, a small number of predictions per class lets learning-to-defer algorithms outperform the AI and expert individually.

  • Problem

    Learning to defer requires individual expert predictions in addition to ground-truth labels, and acquiring enough predictions can be costly, especially when experts change frequently.

    Source: Introduction · p. 1, Related Work · p. 2

  • Method

    The approach trains an embedding model with ground-truth labels, learns an expertise predictor from a small set of expert predictions, and generates predictions for unlabeled instances.

    Source: Introduction · p. 1

  • Results

    With six expert predictions per class, all evaluated algorithms exceeded individual AI and radiologist performance, averaging 98.96% of the accuracy achieved with complete expert predictions.

    Source: Introduction · p. 2, Conclusion · p. 8

  • Takeaways & Limitations

    The evaluation supports using a small number of expert predictions per class to train learning-to-defer algorithms that outperform the individual AI and human expert.

    Source: Conclusion · p. 8

    Artificial expert predictions used in training can risk discriminatory outcomes, so fairness should be monitored in application settings.

    Source: Ethics Statement · p. 8

Abstract

from arXiv · show

Recent research suggests that combining AI models with a human expert can exceed the performance of either alone. The combination of their capabilities is often realized by learning to defer algorithms that enable the AI to learn to decide whether to make a prediction for a particular instance or defer it to the human expert. However, to accurately learn which instances should be deferred to the human expert, a large number of expert predictions that accurately reflect the expert's capabilities are required -- in addition to the ground truth labels needed to train the AI. This requirement shared by many learning to defer algorithms hinders their adoption in scenarios where the responsible expert regularly changes or where acquiring a sufficient number of expert predictions is costly. In this paper, we propose a three-step approach to reduce the number of expert predictions required to train learning to defer algorithms. It encompasses (1) the training of an embedding model with ground truth labels to generate feature representations that serve as a basis for (2) the training of an expertise predictor model to approximate the expert's capabilities. (3) The expertise predictor generates artificial expert predictions for instances not yet labeled by the expert, which are required by the learning to defer algorithms. We evaluate our approach on two public datasets. One with "synthetically" generated human experts and another from the medical domain containing real-world radiologists' predictions. Our experiments show that the approach allows the training of various learning to defer algorithms with a minimal number of human expert predictions. Furthermore, we demonstrate that even a small number of expert predictions per class is sufficient for these algorithms to exceed the performance the AI and the human expert can achieve individually.

Introduction

Learning to defer algorithms can combine human and AI capabilities, but training them requires many individual expert predictions in addition to ground-truth labels. The paper proposes generating artificial predictions from a small labeled subset and evaluates this approach across synthetic and real-world expert settings.

Introduction

  • Learning to defer can combine AI and expert capabilities, but it requires ground-truth labels and additional expert predictions that can be costly to collect.The burden is especially relevant when experts change frequently or specialist labeling is time-consuming.Source: Introduction · p. 1

Approach

  • The approach trains an embedding model on ground-truth labels, fits an expertise predictor using a small set of individual expert predictions, then generates artificial predictions for unlabeled instances.The expert’s available and artificial predictions are used to train learning to defer algorithms.Source: Introduction · pp. 1–2

Introduction

  • With six expert predictions per class, all evaluated algorithms exceeded individual AI and radiologist performance, averaging 98.96% of the accuracy achieved with complete expert predictions.The evaluation uses synthetic experts on CIFAR-100 and real radiologists’ predictions on the NIH chest X-ray dataset.Source: Introduction · p. 2

Related Work

  • Prior learning-to-defer methods use expert predictions to identify strengths and weaknesses; approaches include confidence-based deferral and jointly trained classifiers and deferral systems (Raghu et al. 2019; Madras, Pitassi, and Zemel 2018; Okati, De, and Rodriguez 2021; Wilder, Horvitz, and Kamar 2020).Other work studies theoretical guarantees, surrogate losses, bandit feedback, and allocation across teams of experts (De et al. 2020; De et al. 2021; Mozannar and Sontag 2020; Gao et al. 2021; Hemmer et al. 2022a; Keswani, Lease, and Kenthapadi 2021).Source: Related Work · p. 2
  • Because expert capabilities can vary across the feature space, prior methods require individual predictions for each collaborating expert; this approach instead infers predictions for unlabeled instances without iterative queries.Concurrent work reduces prediction requirements through active learning (Charusaie et al. 2022).Source: Related Work · p. 2

Problem Formulation

  • For classification, the setting includes instances, ground-truth labels, and individual expert predictions, which may differ from ground truth because expert capabilities vary across the feature space.In medical applications, adjudicated labels may come from a panel different from the expert who later collaborates with the AI.Source: Problem Formulation · p. 3
  • The formulation assumes a small subset of instances has individual expert predictions, while the remaining instances have ground-truth labels but no individual expert predictions.The method generates artificial predictions for the latter subset so the deferral system and classifier can be trained on all instances.Source: Problem Formulation · p. 3

Approach

  • The proposed approach is a prior step that can supply artificial expert predictions for training a variety of learning to defer algorithms.Its three stages are embedding training, expertise prediction, and generation of predictions for instances without expert labels.Source: Approach · p. 3

Training of Embedding Model

The embedding model maps each instance to a feature representation, while a classifier learns the ground-truth classification task from those representations.

Training of Embedding Model

  • The embedding model maps each instance into a feature representation using a feature extractor.The model is denoted Φ_emb: X → R^d, where d is the feature-vector length.Source: Training of Embedding Model · p. 3
  • A classifier uses the embedding representations to learn the classification task from all instances and their ground-truth labels.The classifier is denoted Ω: R^d → Y.Source: Training of Embedding Model · p. 3
  • The trained embedding maps each instance to a feature vector that constrains the hypothesis space for learning the expert’s capabilities.The paper describes this as transferring prior knowledge from ground-truth prediction to the task of generating artificial expert predictions.Source: Training of Embedding Model · p. 3

Training of Expertise Predictor Model

The expertise predictor approximates an individual expert’s labeling behavior, using feature representations and a binary target for whether the expert’s prediction is correct.

Training of Expertise Predictor Model

  • For multiclass tasks, the approach simplifies expert-capability learning to predicting whether the expert’s answer is correct or incorrect.The paper motivates this binary target as avoiding the need to learn which incorrect class the expert predicts when errors are not systematic confusions.Source: Training of Expertise Predictor Model · p. 3
  • The approach translates each multiclass expert prediction into a binary expert prediction for the expertise-prediction task.The resulting binary predictions and feature representations are used to train the expertise predictor.Source: Training of Expertise Predictor Model · p. 3
  • An interchangeable semi-supervised learning component incorporates labeled and unlabeled instances when training the expertise predictor.FixMatch and CoMatch instantiate this component to demonstrate its flexibility.Source: Training of Expertise Predictor Model · p. 3

3) Generation of Artificial Expert Predictions

Embedding-FixMatch and Embedding-CoMatch use semi-supervised learning losses to train the expertise predictor from human predictions and unlabeled instances. CoMatch adds a contrastive loss to FixMatch’s supervised and unsupervised terms.

Generation of Artificial Expert Predictions

  • Embedding-FixMatch and Embedding-CoMatch instantiate the approach with semi-supervised methods combining consistency regularization and entropy minimization.FixMatch uses supervised and unsupervised loss terms; CoMatch adds a contrastive loss term.Source: 3) Generation of Artificial Expert… · p. 4
  • The supervised loss compares the model’s prediction for a weakly augmented labeled image with its binary expert prediction.The unsupervised loss uses predictions from weakly augmented unlabeled images to train on strongly augmented versions; FixMatch retains hard pseudo-labels above a threshold, whereas CoMatch uses memory-smoothed pseudo-labels.Source: 3) Generation of Artificial Expert… · p. 4
  • CoMatch additionally aligns label-space and embedding-space similarity graphs for unlabeled instances through a contrastive loss.Its expertise predictor uses a projection head to map feature representations to a low-dimensional embedding.Source: 3) Generation of Artificial Expert… · p. 4

Generation of Artificial Expert Predictions

The expertise predictor applies the embedding model and predictor to unlabeled instances, then translates artificial binary predictions into multiclass predictions. The experiments evaluate the approach on CIFAR-100 and the NIH dataset.

Generation of Artificial Expert Predictions

  • For each unlabeled instance, the embedding model and expertise predictor produce an artificial binary expert prediction, which is then translated into a multiclass prediction.Figure 1 illustrates the approach using FixMatch as an example semi-supervised component.Source: Generation of Artificial Expert… · p. 4

Experiments

  • Experiments evaluate the approach on CIFAR-100 with synthetic experts and NIH with real-world radiologists’ predictions.Further implementation details and results are reported in the Appendix.Source: Experiments · p. 4

Experimental Setup

The evaluation measures both artificial expert prediction quality and the performance of learning-to-defer algorithms trained with artificial and human predictions. It compares these methods against complete-expert-prediction, human-only, and classifier-only benchmarks.

Experimental Setup

  • The evaluation first measures how well artificial binary predictions match the human experts’ true binary predictions, then compares learning-to-defer performance with complete expert-prediction training.The experiments vary the number of available human predictions using l = m*k, with m in {2, 4, 6, 10, 20, 50, 250}.Source: Experimental Setup · pp. 4–5
  • Artificial binary expert predictions are evaluated with F0.5-score because the correct-versus-incorrect prediction classes can be imbalanced.For each l, instances are randomly drawn from training data while maintaining balanced class proportions.Source: Experimental Setup · p. 5
  • Learning-to-defer algorithms are evaluated using system accuracy relative to complete-expert-prediction training, with human-only and classifier-only performance as lower boundaries.The datasets are balanced across classes.Source: Experimental Setup · p. 5
  • The comparison includes Embedding-FixMatch and Embedding-CoMatch, FixMatch and CoMatch without embeddings, and embedding-based supervised NN and SVM predictors.The embedding supervised baselines combine an embedding model with a supervised classifier as the expertise predictor.Source: Experimental Setup · p. 5

CIFAR-100

On CIFAR-100, the experiments model synthetic experts with subclass-specific strengths and weaknesses and test learning-to-defer algorithms using artificial expert predictions. Embedding-SSL approaches yield strong defer performance with few expert predictions and outperform comparison baselines at low label counts.

CIFAR-100

  • Synthetic experts predict correctly on selected strength subclasses, with additional strengths weighted by similarity to a base; on weakness subclasses, predictions are similarity-weighted random choices.The task maps the 100 subclasses to 20 superclasses for classification.Source: CIFAR-100 · pp. 5–6
  • 99.02% (H60) and 98.81% (H90) of the upper boundary are reached with six expert predictions per class under Mozannar and Sontag; all three algorithms beat individual team members.Figure 2 covers Mozannar and Sontag (2020), Raghu et al. (2019), and Okati, De, and Rodriguez (2021) for H60 and H90 on CIFAR-100.Source: Figure 2 · p. 6, CIFAR-100 · p. 6
  • For l < 1,000, proposed Embedding-SSL approaches outperform Embedding-SL baselines on average by 4.09% for H60 and 1.97% for H90.Source: CIFAR-100 · p. 6
  • For l < 400, Embedding-SSL improves SSL baselines by 1.78% (H60) and 1.68% (H90) for Raghu, and 1.96% (H60) and 3.92% (H90) for Okati.For Raghu, improvements over Embedding-SL are 0.53% (H60) and 1.22% (H90); for Okati, they are 0.40% (H60) and 0.06% (H90) over all l.Source: CIFAR-100 · pp. 6–7

NIH Chest X-rays

On the NIH chest X-ray evaluation, Embedding-SSL approaches support learning-to-defer systems trained with few expert predictions, with results varying by algorithm and baseline. The paper concludes that small per-class expert-label budgets can suffice, while emphasizing fairness monitoring and identifying random selection of expert-labeled instances as a limitation of the current approach.

NIH Chest X-rays

  • For l < 100, Embedding-SSL improved artificial-prediction performance over SSL baselines by 38.83% on average and over Embedding-SL baselines by 7.04% on average.For l ≥ 100, Embedding-SL baselines were on par with the proposed approaches.Source: NIH Chest X-rays · p. 7
  • With six expert predictions per class, the Mozannar and Sontag algorithm achieved 99.12% and the Raghu et al. and Okati et al. algorithms achieved 98.83% of their performance when trained with all available expert predictions.The NIH experiment focused on airspace opacity; system accuracies exceeded the lower baselines for all tested expert-prediction counts.Source: NIH Chest X-rays · pp. 7–8
  • For l < 100, Embedding-SSL improved learning-to-defer performance over baselines across the three evaluated algorithms, with reported average gains ranging from 0.69% to 2.42%.The improvements were relative to SSL or Embedding-SL baselines, depending on the algorithm.Source: NIH Chest X-rays · p. 8
  • The evaluation found no evidence that artificial expert predictions systematically amplified bias, although average age-bias differences increased for all three algorithms.Gender-bias differences decreased for Mozannar and Sontag and Raghu et al., but increased for Okati et al.Source: NIH Chest X-rays · p. 8

Conclusion

  • The approach trains an expertise predictor from ground-truth-based embeddings and uses its artificial expert predictions alongside available expert predictions to train learning-to-defer algorithms.The conclusion says the approach accurately learns an individual expert’s capabilities.Source: Conclusion · p. 8
  • The authors identify random selection of instances for expert labeling as a limitation and suggest exploring strategies to select representative instances.They state this could further improve effectiveness and potentially ease onboarding new experts.Source: Conclusion · p. 8

Ethics Statement

  • The authors caution that using artificial predictions with few human predictions can risk discriminatory systems and stress monitoring fairness in application settings.Source: Ethics Statement · p. 8

Related papers

Loading 2304.07306v1…