Introduction
Evaluating the performance of AI models typically relies on human-labeled datasets. However, in real-world deployments, acquiring labels for new data is often slow, expensive, or impossible for real-time monitoring. This creates a significant challenge when models encounter "distribution shift"—situations where the operational data differs from the training data, potentially leading to silent performance degradation.
The field of Automatic Model Evaluation (AutoEval) seeks to predict a model's accuracy on unlabeled test sets autonomously. Current approaches often suffer from global over-simplification, ignoring local variations within the data distribution, or failing to unify model confidence and internal features into a single framework. This paper introduces Divide-and-Aggregate Learning (DAL), a novel framework that addresses these issues by analyzing local distribution shifts to predict performance drops without requiring new labels.
Methodology: Divide-and-Aggregate Learning (DAL)
DAL treats unsupervised evaluation as a supervised meta-learning task. The framework operates through a multi-step process designed to capture both global and local distribution characteristics.
1. Meta-Dataset Generation
To train the evaluator, the authors create a meta-dataset by applying various transformations (e.g., rotations, blurs, color shifts) to a labeled validation set. This generates thousands of synthetic test sets with known ground-truth accuracies, providing the necessary supervision for the regression model.
2. Divide Learning
Rather than analyzing the entire dataset as a single entity, DAL partitions the model's output—either internal features or softmax confidence scores—into $k$ "buckets" using binning strategies like equal-frequency or equal-width. This "divide" step is crucial for capturing local information and identifying specific clusters where the model might be struggling.
3. Aggregate Learning
To quantify the shift between the training and test distributions, DAL calculates a symmetric version of the Kullback-Leibler (KL) divergence ($KL(p||q) + KL(q||p)$) for each bucket. These individual shift measurements are then aggregated into a comprehensive feature vector that represents the total distribution shift.
4. Regression Model
A simple Linear Regression model is trained on the meta-dataset to map these aggregated distribution shifts to the actual observed accuracy. Once trained, this regressor can predict the accuracy of a deployed model on any brand-new, unlabeled dataset based solely on its output distribution shifts.
Experimental Setup and Datasets
The researchers evaluated DAL across four major computer vision benchmarks:
- CIFAR-10: Including variants like CIFAR-10.1, CIFAR-10.1-C (corrupted), and CIFAR-10-F (Flickr).
- MNIST: Handwritten digit classification.
- TinyImageNet: A 200-class subset of ImageNet.
- ImageNet: The industry-standard 1,000-class dataset.
The method was tested across various model architectures, including ResNet, RepVGG, and AlexNet, to ensure generalizability.
Key Results and Insights
DAL consistently outperformed existing state-of-the-art (SOTA) methods such as ConfScore, ATC, and KCFCA. Performance was measured using Root Mean Square Error (RMSE) between predicted and true accuracy.
- CIFAR-10 Performance: For a ResNet-56 model, DAL-f (feature-based) achieved an RMSE of 3.570, a significant improvement over the previous SOTA (KCFCA at 6.766).
- ImageNet Performance: Using a ResNet-18, DAL-f achieved an RMSE of 2.022, the lowest among all tested methods.
- Local Heterogeneity: A key finding is that global metrics often mask local shifts. By dividing data into buckets, DAL can detect failures in specific data sub-types even when the overall distribution appears stable.
- Metric Stability: The use of symmetric KL divergence proved essential for providing stable features to the regression model.
Limitations and Future Work
While DAL shows strong performance, several avenues for future research remain:
- Non-linear Regression: Exploring more complex models like small neural networks for the final prediction might capture more subtle relationships between shift and accuracy.
- Hyperparameter Sensitivity: The choice of binning strategy and the number of buckets ($k$) affects performance; finding an automated way to tune these parameters would be beneficial.
- Cross-Domain Application: Future work should investigate DAL's effectiveness in domains beyond computer vision, such as Natural Language Processing (NLP) or tabular data analysis.
Summary updated