Source-linked AI summary
Human uncertainty makes classification more robust
Joshua C. Peterson, Ruairidh M. Battleday, Thomas L. Griffiths, Olga Russakovsky
TL;DR
CNN accuracy on standard benchmarks is nearing saturation, while generalization beyond training data and adversarial robustness remain limited. The paper collects full human label distributions in CIFAR10H and trains classifiers on these soft labels, finding better out-of-sample generalization, closer human-uncertainty modeling, and greater adversarial resistance.
Problem
CNNs can achieve near-perfect top-1 accuracy yet remain weak on out-of-training-distribution data and adversarial examples.
Method
The paper introduces CIFAR10H, full human label distributions for the CIFAR10 test set, and trains CNNs using these human soft labels.
Results
Training with human soft labels improves out-of-sample generalization, human-uncertainty fit, and resistance to adversarial attacks compared with hard-label controls.
Takeaways & Limitations
Human category uncertainty can help protect classifiers against distributional shift and adversarial attacks while enabling finer-grained evaluation of similarity structure.
Takeaways & Limitations
The data-collection method does not immediately scale to larger training sets, so the dataset is primarily positioned as a testbed for larger-scale algorithms.
Abstract
from arXiv · showhide
The classification performance of deep neural networks has begun to asymptote at near-perfect levels. However, their ability to generalize outside the training set and their robustness to adversarial attacks have not. In this paper, we make progress on this problem by training with full label distributions that reflect human perceptual uncertainty. We first present a new benchmark dataset which we call CIFAR10H, containing a full distribution of human labels for each image of the CIFAR10 test set. We then show that, while contemporary classifiers fail to exhibit human-like uncertainty on their own, explicit training on our dataset closes this gap, supports improved generalization to increasingly out-of-training-distribution test datasets, and confers robustness to adversarial attacks.
1. Introduction
Top-1 accuracy is nearing saturation, but CNNs remain weak on distribution shift and adversarial robustness because hard labels omit informative human uncertainty. The paper introduces CIFAR10H and shows that training on full human label distributions improves generalization, human-uncertainty modeling, and adversarial resistance.
- Motivation: CNNs can match or exceed human top-1 accuracy while performing poorly on related stimuli and adversarial examples.Top-1 accuracy measures only correspondence between the model’s most probable label and the held-out image’s ground-truth label.
- Motivation: Full label distributions preserve informative classification errors and better represent perceptual similarity than modal labels alone.For example, confusing a dog with a cat may indicate better generalization than confusing it with a truck.
- Contributions: CIFAR10H provides full label distributions for all 10,000 CIFAR10 test images using over 500k crowdsourced human categorization judgments.The dataset is intended to measure human uncertainty at image level.
- Contributions: CNNs trained with CIFAR10H soft labels generalize better to out-of-sample datasets than hard-label controls.The benchmark also evaluates how closely models fit human labels and uncertainty.
- Contributions: Training on human labels makes CNNs significantly more resistant to adversarial attacks.The paper frames this as a benefit beyond conventional top-1 performance.
2. Related Work
Related work has used class relationships, soft labels, and human-provided distributions to improve classification or robustness. This paper differs by deriving image-specific relationships from human confusion and providing a large-scale evaluation testbed.
- Hierarchical Classification: Earlier methods used class hierarchies for shared representations, model combination, hierarchical prediction, and hierarchical evaluation.The paper argues that class confusion is increasingly relevant as modern datasets contain overlapping object classes and top-K accuracy improves.
- Knowledge Distillation: This work resembles manual hierarchy construction but infers image-specific class relationships from human confusion rather than explicitly constructing a hierarchy.The comparison connects the paper to knowledge-distillation and hierarchy-based approaches.
- Soft Labels: Soft-label research includes label smoothing, test-time human uncertainty, and mixup’s convex combinations of examples and hard labels.The paper uses human confusion as the source of soft labels rather than an automated interpolation rule.
- Human Studies: Human distributional labels have also been studied in medical diagnosis, but those studies did not provide a large-scale testbed for evaluating classification models.The paper notes that uncertainty labels can sometimes arise automatically during data collection.
3. From Labels to Label Distributions
The paper argues that one-hot modal labels are a biased approximation of the conditional label distribution p(y|x). It instead uses human label distributions to train classifiers against the uncertainty present in image categories.
- Training Objective: The training goal is to minimize expected loss on unobserved labels for future images drawn from the same underlying data distribution.The dataset provides the human distribution directly for training or evaluation against human labeling behavior.
- From Modal Labels to Distributions: One-hot labels are optimal only when each image has zero probability for every category except the modal human choice.Human and network confusions show that this assumption can fail for individual stimuli.
- From Modal Labels to Distributions: The proposed objective trains classifiers to capture the full distribution over labels rather than only the most likely category.The approach treats errors as informative when they reflect meaningful category confusions.
- Training Objective: For probabilistic classifiers with negative log-likelihood loss, minimizing expected loss under human labels becomes cross-entropy between human and predicted label distributions.The human distribution phum(y|x) serves as an estimator of the underlying category distribution for an image.
4. Dataset Construction
The authors construct CIFAR10H by collecting human judgments for every image in the CIFAR10 test set. The resulting benchmark provides substantial distributional label data across all ten categories.
- Dataset Choice: CIFAR10 was selected because it is actively used for state-of-the-art classification, small enough for comprehensive human data collection, and low-resolution enough to elicit varied responses.These properties make the full test set practical for measuring human uncertainty.
- Stimulus Set: CIFAR10H contains human judgments for all 10,000 32 × 32 color images in the CIFAR10 testing subset.The images comprise 1,000 examples from each of ten categories, enabling evaluation against a distribution over labels rather than a single label.
- Human Data Collection: The study collected 511,400 classifications from 2,571 Amazon Mechanical Turk participants.Each participant categorized 200 images, with shuffled label positions and recurring attention checks after training.
5. Generalization Under Distributional Shift
CNNs trained on CIFAR10H human soft labels are evaluated against one-hot controls across datasets with increasing distributional shift. Soft-label training improves accuracy, crossentropy, SBA, and human-like uncertainty behavior, with larger gains on stronger shifts.
- Experimental Setup: The experiments train eight CNN architectures with CIFAR10H soft labels and compare them with one-hot controls across increasingly out-of-sample datasets.Models are evaluated using accuracy and crossentropy, while SBA is reported for held-out CIFAR10H folds.
- Generalization Results: Soft-label CNNs improve accuracy on every generalization dataset relative to one-hot controls, with the boost increasing as test distributions move farther out of training distribution.The comparison spans standard CIFAR10, near-sample datasets, CINIC10, and ImageNet-Far.
- Generalization Results: Accuracy rises from 49.4% to 51.4% on ImageNet-Far with human soft labels, compared with a 1% improvement from 83.5% to 84.5% on CIFAR10.The ImageNet-Far result represents a 2% average gain under stronger distributional shift.
- Confidence and Calibration: Crossentropy falls from 2.9 to 1.8 on ImageNet-Far, a 38% reduction, compared with a 29% reduction from 0.7 to 0.5 on CIFAR10.The authors interpret this as better confidence in correct choices and more probability assigned to the ground truth during errors.
- Confidence and Calibration: Soft-label CNNs perform on average 5% better than controls on second best accuracy across CIFAR10H held-out folds.SBA evaluates the broader quality of the two most likely categories, including the nature of incorrect guesses.
- Confidence and Calibration: Soft-label models are far less confident when incorrect, only marginally less confident when correct, and better fit human uncertainty patterns.Their predictions distribute probability mass more like people while retaining the same top choice.
6. Alternative Soft Label Methods
The authors compare human image-level soft labels with simpler alternatives, including hard-label fine-tuning, class-level penalties, knowledge distillation, mixup, and label sampling. Human targets provide the best fit to human uncertainty and the strongest reported generalization among the compared fine-tuning approaches.
- Results: Fine-tuning on human targets produces the best CIFAR10.1 generalization in crossentropy, while Table 1 evaluates human-label fit and further generalization.The reported holdouts are CIFAR10.1v4 and CIFAR10.1v6.
- Methods: Human-label training is compared against hard-label controls, class-level penalties, knowledge distillation, mixup, and sampling human labels.The evaluation uses crossentropy on human labels, ground-truth labels, and CIFAR10.1v4 and CIFAR10.1v6 holdouts.
- Results: Pretrained models fit ground-truth labels better than human soft labels, indicating that hard-label training does not infer the human label distributions.An eight-model ensemble improves human fit but remains less similar to human supervision.
- Results: Fine-tuning improves human fit, with image-level human soft labels and sampling from them performing best among the tested methods.Category-level soft labels also help, but less than the image-level human targets.
7. Robustness to Adversarial Attacks
The paper evaluates adversarial robustness using FGSM and PGD on models trained with hard versus human soft labels. Human soft-label training lowers attack crossentropy and slows PGD-driven loss growth, with modest accuracy benefits under FGSM and a small residual accuracy under PGD.
- FGSM: FGSM attacks produce much lower crossentropy after human-label fine-tuning than after hard-label fine-tuning.The reduction is roughly half across the evaluated architectures, and no additional defensive training is required.
- FGSM: For five of eight architectures, FGSM accuracy also improves with human soft targets, including the largest differences for Wide ResNet and ResNet preact.Table 2 reports accuracy and crossentropy after attacks on CIFAR10- and CIFAR10H-tuned networks.
- PGD: Under PGD, accuracy falls to 0% for standard-label networks but remains 1% for networks trained with human labels.These results are obtained without active defensive training.
- PGD: PGD increases loss more slowly after soft-label fine-tuning, so successfully attacking human-like networks requires substantially more effort.Figure 4 plots crossentropy against PGD iteration and shows faster loss growth for standard-label models.
8. Discussion
The discussion argues that image-level human category uncertainty can help protect classifiers against distributional shift and adversarial attacks while enabling finer-grained evaluation of learned similarity structure. The dataset is presented as an initial benchmark for measuring progress toward human-like robustness, although its collection method may not scale directly to larger training sets.
- Implications: Image-level human category uncertainty helps protect classifiers against distributional shift and adversarial attacks.The authors contrast this with common classification benchmarks, which do not naturally provide such protections.
- Implications: CIFAR10H enables evaluation of whether learning algorithms infer useful similarity structure beyond top-1 performance.The dataset can serve as a testing environment even when it is not used for training.
- Limitation: The data-collection method does not immediately appear scalable to larger training sets.The authors suggest mining labels from approximately 30% of images with concentrated human uncertainty as a more efficient alternative.