Rate-In: Information-Driven Adaptive Dropout Rates for Improved Inference-Time Uncertainty Estimation
Abstract
Rate-In measures the information loss introduced by dropout and adapts its rate for each layer and input. Experiments on synthetic data and medical imaging tasks examine calibration and predictive uncertainty relative to fixed dropout settings.
Type
Publication
IEEE/CVF Conference on Computer Vision and Pattern Recognition
Ravid Shwartz-Ziv
AI Researcher
AI researcher at Meta MSL with a background in information theory and computational neuroscience, working on world models, memory, and compression. Former Assistant Professor and Faculty Fellow at NYU, collaborating on research across academia and industry.