A Preliminary Study to Examining Per-class Performance Bias via Robustness Distributions
Bibliographic record
Abstract
As neural networks are increasingly used in sensitive real-world applications, mitigating bias of classifiers is of crucial importance. One often-used approach to controlling quality in classification tasks is to ensure that predictive performance is balanced between different classes; however, it has been shown in previous work that even if class performance is balanced, instances of some classes are easier to perturb in such a way that they are misclassified, which indicates that per-class performance bias exists. In this preliminary study, we found that even when class performance is balanced, class robustness can vary strongly when assessing the robustness of a given neural network classifier in a more nuanced fashion. For this purpose, we use robustness distributions, i.e., empirical probability distributions of some robustness metric, such as the critical epsilon value, over a set of instances. We observed that the robustness of the same class over the same data can significantly differ from each other for different neural networks; this means that even when a neural network appears to be unbiased, it might be easier to perturb instances of a given class so that they are misclassified. Furthermore, we explored the robustness distributions when we have a predefined target class, i.e., a specific class into which an instance is misclassified after perturbation. Our empirical results indicate that in most cases, there are significant differences in robustness distributions for different classes. While our empirical results reported here are for MNIST classifiers, we are currently performing experiments using the German Traffic Sign Recognition Benchmark. Furthermore, we are running experiments with retrained networks for fairness, to see whether this has a significant effect on the per-class robustness distributions. Lastly, we aim to create a robust class fairness metric based on our findings.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.000 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.001 | 0.002 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.006 | 0.004 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".