Paper Image

Predicting the most probable image labels

Published on:

28 March 2024

Primary Category:

Machine Learning

Paper Authors:

Anqi Mao,

Mehryar Mohri,

Yutao Zhong

Bullets

Key Details

Common multi-class losses work for top-k classification

New cardinality-aware losses balance accuracy and k

Algorithms using these losses adapt k per image

Experiments show higher accuracy for given k

AI generated summary

Predicting the most probable image labels

This paper studies top-k classification, which predicts the k most probable labels for an image, going beyond single-label prediction. It shows that common multi-class loss functions like cross-entropy admit consistency guarantees for top-k classification. To balance accuracy and k, it introduces cardinality-aware losses optimized by new algorithms that adaptively choose k per image. Experiments on image datasets demonstrate these algorithms achieve higher accuracy for a given k.

Answers from this paper

Comments

No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up