Online Platt Scaling with Calibeating
Jul 24, 2023
Speakers
Chirag Gupta
Speaker · 0 followers
Aaditya Ramdas
Speaker · 3 followers
About
We develop online, adaptive methods for post-hoc calibration. First, we propose Online Platt Scaling (OPS), an online version of the popular Platt scaling technique. OPS is based on online logistic regression, a well-studied problem in online learning. We instantiate OPS with the Online Newton Step algorithm that has logarithmic regret and linear running time, but any other alternative may be plugged in. Second, for scenarios where the best Platt scaling model is itself miscalibrated, we make OPS more robust by combining it with a recently developed technique called calibeating. The resulting OPS + calibeating method is guaranteed to be calibrated in both i.i.d. and adversarial settings, and performs the best (without tuning) on a variety of synthetic and real-world datasets.We develop online, adaptive methods for post-hoc calibration. First, we propose Online Platt Scaling (OPS), an online version of the popular Platt scaling technique. OPS is based on online logistic regression, a well-studied problem in online learning. We instantiate OPS with the Online Newton Step algorithm that has logarithmic regret and linear running time, but any other alternative may be plugged in. Second, for scenarios where the best Platt scaling model is itself miscalibrated, we make OP…
Like the format? Trust SlidesLive to capture your next event!
Professional recording and live streaming, delivered globally.
Sharing
Recommended Videos
Presentations on similar topic, category or speaker
Differentially Private Distributed Bayesian Linear Regression with MCMC
Watch later
Guiding Pretraining in Reinforcement Learning with Large Language Models
Watch later
Yuqing Du, …
Three Spotlight Talks
Watch later
A Conditional Normalizing Flow for Accelerated Multi-Coil MR Imaging
Watch later
Jeffrey Wen, …
Conformal Prediction Sets for Graph Neural Networks
Watch later
