ACLS: Adaptive and Conditional Label Smoothing for Network Calibration

TL;DR

This paper introduces ACLS, a novel loss function for network calibration that unifies and improves upon existing regularization-based label smoothing methods, with extensive experiments showing its effectiveness.

Contribution

We analyze existing regularization methods for calibration, revealing their relation to label smoothing, and propose ACLS, a new loss that combines their strengths while addressing limitations.

Findings

ACLS improves calibration accuracy across benchmarks

Regularization methods relate to label smoothing principles

Extensive experiments validate ACLS's effectiveness

Abstract

We address the problem of network calibration adjusting miscalibrated confidences of deep neural networks. Many approaches to network calibration adopt a regularization-based method that exploits a regularization term to smooth the miscalibrated confidences. Although these approaches have shown the effectiveness on calibrating the networks, there is still a lack of understanding on the underlying principles of regularization in terms of network calibration. We present in this paper an in-depth analysis of existing regularization-based methods, providing a better understanding on how they affect to network calibration. Specifically, we have observed that 1) the regularization-based methods can be interpreted as variants of label smoothing, and 2) they do not always behave desirably. Based on the analysis, we introduce a novel loss function, dubbed ACLS, that unifies the merits of existing regularization methods, while avoiding the limitations. We show extensive experimental results for image classification and semantic segmentation on standard benchmarks, including CIFAR10, Tiny-ImageNet, ImageNet, and PASCAL VOC, demonstrating the effectiveness of our loss function.