Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Feature Clipping for Uncertainty Calibration

About

Deep neural networks (DNNs) have achieved significant success across various tasks, but ensuring reliable uncertainty estimates, known as model calibration, is crucial for their safe and effective deployment. Modern DNNs often suffer from overconfidence, leading to miscalibration. We propose a novel post-hoc calibration method called feature clipping (FC) to address this issue. FC involves clipping feature values to a specified threshold, effectively increasing entropy in high calibration error samples while maintaining the information in low calibration error samples. This process reduces the overconfidence in predictions, improving the overall calibration of the model. Our extensive experiments on datasets such as CIFAR-10, CIFAR-100, and ImageNet, and models including CNNs and transformers, demonstrate that FC consistently enhances calibration performance. Additionally, we provide a theoretical analysis that validates the effectiveness of our method. As the first calibration technique based on feature modification, feature clipping offers a novel approach to improving model calibration, showing significant improvements over both post-hoc and train-time calibration methods and pioneering a new avenue for feature-based model calibration.

Linwei Tao, Minjing Dong, Chang Xu• 2024

Related benchmarks

TaskDatasetResultRank
Image ClassificationImageNet LT--
264
Image ClassificationCIFAR-100-LT mild (R-10) (test)
ECE3.19
35
Image ClassificationCIFAR-100
ECE1.7
35
Image ClassificationCIFAR-100-LT extreme (R-100) (test)
ECE2.37
35
Medical Image ClassificationMedical Image Classification 10 datasets avg
ECE3.91
35
Image ClassificationCIFAR-10
ECE0.5
35
Classification20News
Error Rate37.1
32
ClassificationTImageNet
Error Rate41.69
32
ClassificationImageNet-1K
Error Rate18.94
16
Image ClassificationImageNet-1K
Error Rate18.94
16
Showing 10 of 18 rows

Other info

Follow for update