Selective classification to enable reliable deep learning deployment

Paper Title:

Selective Classification Under Distribution Shifts

Published on:

8 May 2024

Primary Category:

Machine Learning

Paper Authors:

Hengyue Liang,

Le Peng,

Ju Sun

Bullets

Key Details

•

Proposes generalized selective classification to handle multiple data shifts

•

Identifies scale sensitivity issues of softmax-based confidence scores

•

Develops two effective margin-based confidence scores

•

Shows superior reliability of margins for selectivity

•

Analyzes favorable rejection patterns of margin scores

Explore the topics in this paper

classification

deep learning

distribution shift

model prediction

rejection technique

AI generated summary

Selective classification to enable reliable deep learning deployment

This paper proposes a selective classification framework that allows for distribution shifts between training and deployment. It focuses on developing non-training-based confidence scores to reject likely-incorrect predictions on deep learning models. The key ideas are margin-based scores that are more reliable than existing methods under label/covariance shifts.