Taha Belkhouja

When the data collected is distorted, is the commonly used Minkowski distance appropriate to compare the similarity of the signals? Time-series perturbations such as time-shifts and frequency-distortion affect significantly the ability of the ML model to classify correctly the input. How can we enhance the robustness of these classifiers using adversarial perturbations and appropriate similarity measures between time-series instances?
Related papers: TSA-STAT (JAIR'22) & DTW-AR (TPAMI'22)

Time-series data acquisition is prone to natural perturbations (such as hardware orientation or amplified noise) that affect significantly the data. Such distortion yields signals that are dissimilar according to Minkowski distances. On the other hand, elastic measures are very expensive to compute. Therefore, how can we enhance the robustness of deep models during training to recognize correctly the distorted inputs?
Related papers: RO-TS (AAAI'22) & StatOpt (ICCAD'22)

One of the failure scenarios in the AI safety domain is confident predictions on Out-Of-Distribution (OOD) examples, essentially for real-world contexts with low tolerance for error. Such examples, not observed during the training phase or outside the intended context of deployment, pose a significant risk of leading to unsafe decision-making outcomes. OOD detectors have the potential of achieving AI systems capable of functioning reliably when presented with these unforeseen examples and improves the predictive uncertainty of deep-learning models.
Related papers: SR Score (TIST'23)

Most practices often train models on specific data and distribute them as black-box models. A significant concern in this practice is the difficulty of rigorously quantifying the uncertainty of black-box algorithms and capturing the deviation of the prediction from the ground truth. Safe deployment of deep neural networks in high-stake real-world applications requires theoretically-sound uncertainty quantification. We study Conformal Prediction for uncertainty quantification of deep models to obtain effective human-ML collaborative systems. We particularly focus on overcoming challenges like imbalanced data distributions, data heterogeneity, distribution shift, and coverage for sub-populations.
Related papers: NCP (AAAI'23) & aPRCP (UAI'23)