Coen Adler

Evan Kellener, Ihina Nath, An Ngo et al.

The introduction of saliency map algorithms as an approach for assessing the interoperability of images has allowed for a deeper understanding of current black-box models with Artificial Intelligence. Their rise in popularity has led to these algorithms being applied in multiple fields, including medical imaging. With a classification task as important as those in the medical domain, a need for rigorous testing of their capabilities arises. Current works examine capabilities through assessing the localization of saliency maps upon medical abnormalities within an image, through comparisons with human annotations. We propose utilizing Segment Anything Model (SAM) to both further the accuracy of such existing metrics, while also generalizing beyond the need for human annotations. Our results show both high degrees of similarity to existing metrics while also highlighting the capabilities of this methodology to beyond human-annotation. Furthermore, we explore the applications (and challenges) of SAM within the medical domain, including image pre-processing before segmenting, natural language proposals to SAM in the form of CLIP-SAM, and SAM accuracy across multiple medical imaging datasets.

Coen Adler, Yuxin Chang, Felix Draxler et al.

The recent development of foundation models for time series data has generated considerable interest in using such models across a variety of applications. Although foundation models achieve state-of-the-art predictive performance, their calibration properties remain relatively underexplored, despite the fact that calibration can be critical for many practical applications. In this paper, we investigate the calibration-related properties of five recent time series foundation models and two competitive baselines. We perform a series of systematic evaluations assessing model calibration (i.e., over- or under-confidence), effects of varying prediction heads, and calibration under long-term autoregressive forecasting. We find that time series foundation models are consistently better calibrated than baseline models and tend not to be either systematically over- or under-confident, in contrast to the overconfidence often seen in other deep learning models.