Search papers, labs, and topics across Lattice.
This study investigates the effectiveness of Test-Time Augmentation (TTA) techniques on tabular-to-image classifiers when faced with distribution shifts. By evaluating six encoding methods and twenty-five TTA strategies across two datasets, the research reveals that TTA significantly enhances out-of-distribution performance, particularly with composite and photometric strategies. Notably, frequency-domain transformations were found to degrade performance, underscoring the importance of selecting appropriate augmentation methods for robustness.
TTA can dramatically boost the robustness of tabular-to-image classifiers under distribution shifts, with certain strategies outperforming others by a significant margin.
Tabular-to-image methods that convert tabular data into visual representations have emerged as a novel paradigm for leveraging the high performance of deep learning models. Despite their advantages, the robustness of these methods under distribution shifts remains under explored. Test-Time Augmentation (TTA) is an effective approach in image classification to improve model generalization and robustness, where predictions over multiple transformed views of each input are aggregated. This work evaluates the impact of TTA techniques on predictive performance under Out-Of-Distribution (OOD) for representations generated by tabular-to-image methods. Six tabular-to-image encoding methods were considered: TINTO, IGTD, DeepInsight, BIE, DistanceMatrix, Fotomics. Twenty-five TTA techniques were used, organized into six types: Geometric, Photometric, Structural, Frequency/Encoding, Mixup, and Composite. We employed two datasets from the TableShift benchmark (HELOC and Voting) that provide in-distribution and OOD test subsets designed to evaluate the effect of distribution shifts on tabular data. The results indicate that TTA improves OOD performance, with composite and photometric strategies providing the best trade-off between robustness and variance. In contrast, frequency-domain transformations that alter the encoder's feature-to-intensity mapping consistently degrade performance. These findings highlight TTA as a promising approach for improving the robustness and generalization of classifiers trained on image representations derived from tabular data, particularly under distribution shifts.
