Colloquia and Seminars

Distribution testing has long been studied in the context of theoretical computer science and statistics. The classical model of distribution testing can be seen as too weak in practice, with even the simplest tasks requiring polynomial (though sublinear) sample complexity. We explore non-classical models tailored to cope with this model's infeasibility for distributions defined over very large data sets. We consider two kinds of models: models that are stronger than the classical one, in which testing can be done more efficiently, and a weaker model, which is better suited for extremely high-dimensional data sets.

The stronger model is the conditional sampling model, in which the algorithm can sample the input distribution when conditioned on subsets of the domain. We consider this model as well as a few of its restricted variants, such as the subcube conditional model. We tighten the bounds for a few core algorithmic tasks in these conditional models.

Finally, we explore the behavior of algorithms in the Huge Object model, which combines the classical distribution testing model with the string testing model to more realistically handle distributions over extremely high-dimensional data sets.

Multi-objective search (MOS) has become essential in robotics, as real-world robotic systems need to simultaneously balance multiple, often conflicting objectives. Recent works explore complex interactions between objectives, leading to problem formulations that do not allow the usage of out-of-the-box state-of-the-art MOS algorithms. In this paper, we suggest a generalized problem formulation that optimizes solution objectives via aggregation functions of hidden (search) objectives. We show that our formulation supports the application of standard MOS algorithms, necessitating only to properly extend several core operations to reflect the specific aggregation functions employed.

We demonstrate our approach in several diverse robotics planning problems, spanning motion-planning for navigation, manipulation and planning for medical systems under obstacle uncertainty as well as inspection planning, and route planning with different road types. We solve the problems using state-of-the-art MOS algorithms after properly extending their core operations, and provide empirical evidence that they outperform by orders of magnitude the vanilla versions of the algorithms applied to the same problems but without objective aggregation.

Conventional unsupervised domain adaptation (UDA) typically relies on updating the predictor's weights to optimize performance on the target domain using a labeled source dataset and unlabeled target data. However, a growing demand for secure deployments, such as closed inference APIs, firmware, or regulatory certification, means models are increasingly released with locked weights. We study domain adaptation in this locked-predictor regime for multivariate time-series models, where the deploying site holds its own labels but source training data may be inaccessible. We introduce PILOT (Per-timestep Input-space adapter for LOcked predicTors), a per-timestep input-space adapter trained against the frozen predictor by back-propagating the task loss using target labels. The framework supports two settings: a source-free variant for strict privacy constraints and a source-assisted variant when data sharing is permitted. We evaluate PILOT across five clinical time-series tasks (encompassing two domain shifts) and five AdaTime sensor benchmarks using a frozen ID-CNN backbone. PILOT achieves state-of-the-art performance across all comparable frozen-backbone, end-to-end, and test-time adaptation baselines. It yields a +15.3 improvement in AUCPR for AKI, increases the mean Macro-F1 score by +10.1 over the leading test-time adaptation method on AdaTime and matches or exceeds the performance of natively trained target-domain models. Furthermore, training PILOT on a single architecture enables zero-shot transfer to other frozen predictors, allowing the same PILOT module to be reused across different architectures.

Cardiac MRI is clinically valuable but inherently slow, requiring many sequential measurements per frame to build a complete image. In dynamic MRI, where each time-frame must be acquired separately, this becomes especially limiting. This work presents a pipeline built on top of TEAM-PILOT model that learns to interpolate videos directly in the frequency domain, generating phase-consistent intermediate frames and effectively enlarging the training dataset without any new acquisitions. The approach addresses two challenges simultaneously: accelerating scan time and alleviating data scarcity, which is a general bottleneck in medical imaging deep learning. We demonstrate that combining 2-shot acquisition (proportional to the amount of signals) with 4 times temporal densification matches standard 8-shot reconstruction quality across multiple state-of-the-art architectures, achieving a 4 times reduction in scan time with no significant loss in image quality.