Enhancing Cross-Architecture Robustness

Background: Dataset Distillation (DD) methods often suffer from architectural overfitting, where the distilled data overfits to the inductive biases of the specific model used during the distillation process.

Question / Future Work: Exploring methods to further enhance the model-agnostic nature of the distilled time series data beyond Harmonic Matching, potentially by incorporating loss terms that explicitly encourage robustness against different model families (e.g., linear models, Transformers, CNNs) during the distillation phase.