Identification and appropriate handling of inconsistencies in data at deployment time is crucial to using machine learning models reliably. Prior works have focused on out-of-distribution data. Instead, with data monitoring at test time, we develop ML methods to characterise incongruous and inconsistent in-distribution (ID) data regions, which may arise from feature space heterogeneity. Identifying such samples can guide understanding of why a model might fail on a specific data point and also guide future data collection. We have addressed this both from a representation learning, as well as, a causal structure perspective.
View posts about…
AI AISTATS Augmented MD AutoML Award Cambridge Adjutorium Cancer causal-deep-learning Causal inference Clinical trials Clinicians Clustering & phenotyping Conference COVID-19 creativity Data-Centric AI Deep learning Early diagnosis genomics ICLR ICML Interpretability ITE inference Journal machine learning Missing data imputation ML ML tools Model neural-differential-equations NeurIPS Organ transplantation Paper Personalized medicine Policy Impact Predictor Quantitative epistemology Reinforcement learning Research team revolutionizing healthcare Software Statistics Synthetic data Time series analysis Tutorial Understanding decision-making
Upcoming events
No event found!
