We demonstrate that cross-frequency interactions, rather than standard marginal spectral statistics, are essential for characterizing and distinguishing the learned dynamics of deep generative models.
We analyze attribution robustness in CNNs and ViTs, showing that SGD can improve robustness of gradient-based attribution, while attention-based attribution requires kernelized attention for such gains.
The paper proposes a spectral framework to analyze the smoothness–faithfulness trade-off in ReLU network explanations, formalizes the resulting explanation gap, and provides principled regularization with empirical validation.
The position paper argues that spectral analysis provides a principled theoretical foundation for understanding and navigating the ante-hoc versus post-hoc trade-offs in explainable AI.
The paper introduces a probabilistic—spectral framework for analyzing deep network explanations, exposes gradient-induced spectral bias, and proposes principled remedies validated empirically.
A DFT-informed multi-model ML framework accurately predicts and rediscovers optimal methane adsorption on functionalized graphene surfaces while enabling sensitivity analysis and manifold-based screening.
 Techniques/card.png)
DeeP4med is an autoencoder architecture that performs domain translation between healthy and cancerous gene expression data to identify the key features that drive disease progression.
We introduce a model-agnostic curriculum learning framework for depth estimation that matches state-of-the-art accuracy while reducing inference time by 71%, enabling fast and energy-efficient deployment on resource-constrained autonomous devices.
We extend a probabilistic heteroscedastic label noise model from regression to classification, utilizing loss attenuation to improve robustness against noisy labels.
