Saved in:
Bibliographic Details
Main Authors: Tang, Yung-Chen, Chen, Pin-Yu, Ho, Tsung-Yi
Format: Preprint
Published: 2022
Subjects:
Online Access:https://arxiv.org/abs/2209.11604
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866929435844804608
author Tang, Yung-Chen
Chen, Pin-Yu
Ho, Tsung-Yi
author_facet Tang, Yung-Chen
Chen, Pin-Yu
Ho, Tsung-Yi
contents Neural network calibration is an essential task in deep learning to ensure consistency between the confidence of model prediction and the true correctness likelihood. In this paper, we propose a new post-processing calibration method called Neural Clamping, which employs a simple joint input-output transformation on a pre-trained classifier via a learnable universal input perturbation and an output temperature scaling parameter. Moreover, we provide theoretical explanations on why Neural Clamping is provably better than temperature scaling. Evaluated on BloodMNIST, CIFAR-100, and ImageNet image recognition datasets and a variety of deep neural network models, our empirical results show that Neural Clamping significantly outperforms state-of-the-art post-processing calibration methods. The code is available at github.com/yungchentang/NCToolkit, and the demo is available at huggingface.co/spaces/TrustSafeAI/NCTV.
format Preprint
id arxiv_https___arxiv_org_abs_2209_11604
institution arXiv
publishDate 2022
record_format arxiv
spellingShingle Neural Clamping: Joint Input Perturbation and Temperature Scaling for Neural Network Calibration
Tang, Yung-Chen
Chen, Pin-Yu
Ho, Tsung-Yi
Machine Learning
Neural network calibration is an essential task in deep learning to ensure consistency between the confidence of model prediction and the true correctness likelihood. In this paper, we propose a new post-processing calibration method called Neural Clamping, which employs a simple joint input-output transformation on a pre-trained classifier via a learnable universal input perturbation and an output temperature scaling parameter. Moreover, we provide theoretical explanations on why Neural Clamping is provably better than temperature scaling. Evaluated on BloodMNIST, CIFAR-100, and ImageNet image recognition datasets and a variety of deep neural network models, our empirical results show that Neural Clamping significantly outperforms state-of-the-art post-processing calibration methods. The code is available at github.com/yungchentang/NCToolkit, and the demo is available at huggingface.co/spaces/TrustSafeAI/NCTV.
title Neural Clamping: Joint Input Perturbation and Temperature Scaling for Neural Network Calibration
topic Machine Learning
url https://arxiv.org/abs/2209.11604