Bridging PCA and Neural Networks: New Insights into Class Bias

Pawel Pukowski, Venet Osmani, Haiping Lu

Keywords: hardness, class bias, class-level hardness identifiers, instance-level hardness identifiers

Abstract:

Understanding class-level hardness is essential for addressing class bias in machine learning. Traditionally, class bias has been explored with two primary approaches: analyzing raw input data to improve preprocessing strategies or examining neural network latent representations to refine model training. In this work, we find that PCA-transformed spaces—despite being produced through linear transformations—still contain substantial information about class-level hardness. This suggests that, despite their distinct goals and methodologies, both PCA and neural networks may encode similar features related to class bias, offering new insights into the nature of class bias and how data representations are formed in both PCA and neural networks.

Analyzing class bias commonly involves Pearson Correlation, which assumes stable inputs. However, we find that class bias is a highly unstable phenomenon with respect to variables such as training time and model initialization, with class-level variability often exceeding the differences between classes. Together with increased variability in class accuracies over dataset-level ones, this suggests that current methods for addressing dataset-level variability may be inadequate for handling class bias.

Primary Area: interpretability and explainable AI

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 13926