[Paper Review] Are Labels Always Necessary for Classifier Accuracy Evaluation?

Weijian Deng, Liang Zheng|arXiv (Cornell University)|Jul 6, 2020

Advanced Neural Network Applications4 citations

TL;DR

This paper proposes AutoEval, a method to estimate classifier accuracy on unlabeled test sets by leveraging a meta-dataset of synthetically transformed training images. Using distribution statistics (mean and covariance) of these transformed datasets as input, a regression model predicts test accuracy with a Spearman correlation of -0.9 between distribution shift and performance, enabling label-free evaluation in real-world deployment scenarios.

ABSTRACT

To calculate the model accuracy on a computer vision task, e.g., object recognition, we usually require a test set composing of test samples and their ground truth labels. Whilst standard usage cases satisfy this requirement, many real-world scenarios involve unlabeled test data, rendering common model evaluation methods infeasible. We investigate this important and under-explored problem, Automatic model Evaluation (AutoEval). Specifically, given a labeled training set and a classifier, we aim to estimate the classification accuracy on unlabeled test datasets. We construct a meta-dataset: a dataset comprised of datasets generated from the original images via various transformations such as rotation, background substitution, foreground scaling, etc. As the classification accuracy of the model on each sample (dataset) is known from the original dataset labels, our task can be solved via regression. Using the feature statistics to represent the distribution of a sample dataset, we can train regression models (e.g., a regression neural network) to predict model performance. Using synthetic meta-dataset and real-world datasets in training and testing, respectively, we report a reasonable and promising prediction of the model accuracy. We also provide insights into the application scope, limitation, and potential future direction of AutoEval.

Motivation & Objective

• Address the challenge of evaluating classifier performance on unlabeled real-world test data where ground-truth labels are unavailable or too costly to obtain.
• Investigate whether model accuracy can be estimated without test labels by modeling the relationship between dataset distribution shifts and classifier performance.
• Develop a meta-dataset of transformed image sets to train a regression model that predicts accuracy based on overall dataset statistics.
• Enable practical deployment monitoring by predicting system reliability in unseen environments without manual annotation.

Proposed method

• Construct a meta-dataset by applying diverse image transformations—such as rotation, background substitution, and foreground scaling—to original training images, creating multiple synthetic datasets.
• Extract global feature statistics (mean and covariance) from the feature representations of each synthetic dataset to represent its distribution.
• Train a regression model (e.g., a neural network) to predict the true classification accuracy of a given classifier on each synthetic dataset using these distribution statistics as input.
• Leverage the strong negative correlation (Spearman’s rho = -0.9) between distribution shift (measured via Fréchet Distance) and model accuracy to guide the learning process.
• Use the trained regression model to predict accuracy on real-world, unlabeled test sets by computing their feature statistics and feeding them into the model.
• Validate the approach using both synthetic meta-datasets and real-world test sets, demonstrating generalization to unseen data distributions.

Experimental results

Research questions

• RQ1 Can classifier accuracy be reliably predicted on unlabeled test sets using only distribution-level statistics of the data?
• RQ2 How strong is the relationship between dataset distribution shift and model performance in real-world deployment scenarios?
• RQ3 To what extent can a meta-dataset of transformed training images generalize to predict accuracy on real-world, unlabeled test sets?
• RQ4 What are the limitations of this approach when faced with out-of-distribution or adversarial test data not covered by the meta-dataset?

Key findings

• A strong negative correlation (Spearman’s rank correlation of -0.9) exists between the Fréchet Distance (FD) measuring distribution shift and the model’s test accuracy, indicating that distribution shift is a reliable proxy for performance degradation.
• The proposed AutoEval framework achieves reasonable and promising accuracy predictions on real-world unlabeled test sets using only feature statistics from the test data, without requiring labels.
• The synthetic meta-dataset effectively covers a wide range of data distributions, enabling generalization to novel real-world test environments.
• The method enables label-free evaluation of model performance, which is critical for real-world deployment monitoring where labeling is impractical or too costly.
• Limitations arise when test sets contain entirely new classes or extreme adversarial conditions not present in the meta-dataset, leading to inaccurate predictions.
• Out-of-distribution detection techniques could potentially be integrated to identify and reject such problematic test sets, improving robustness.