Computer Science · cs.LG
On Calibration of Modern Neural Networks
Chuan Guo, Geoff Pleiss, Yu Sun, Kilian Q. Weinberger
See on the mapAbstract
Confidence calibration -- the problem of predicting probability estimates representative of the true correctness likelihood -- is important for classification models in many applications. We discover that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, we observe that depth, width, weight decay, and Batch Normalization are important factors influencing calibration. We evaluate the performance of various post-processing calibration methods on state-of-the-art architectures with image and document classification datasets. Our analysis and experiments not only offer insights into neural network learning, but also provide a simple and straightforward recipe for practical settings: on most datasets, temperature scaling -- a single-parameter variant of Platt Scaling -- is surprisingly effective at calibrating predictions.
References
The reference list for this paper has not been collected yet.
Cited by 915
915 papers in the index cite this (Semantic Scholar counts 10,157 in total).
The cited-by list is still filling in while reference lists are collected.