Expectation consistency for calibration of neural networks (code)
JSON Export
{
"id": "1892",
"updated": "2023-09-19T14:35:06.004099+00:00",
"metadata": {
"version": 1,
"contributors": [
{
"givennames": "Lucas",
"affiliations": [
"\u00c9cole Polytechnique F\u00e9d\u00e9rale de Lausanne (EPFL), Statistical Physics of Computation lab., CH-1015 Lausanne, Switzerland"
],
"email": "lucas.clarte@epfl.ch",
"familyname": "Clarte"
},
{
"givennames": "Bruno",
"affiliations": [
"D\u00e9partement d\u2019Informatique, \u00c9cole Normale Sup\u00e9rieure - PSL & CNRS, Paris, France"
],
"familyname": "Loureiro"
},
{
"givennames": "Florent",
"affiliations": [
"\u00c9cole Polytechnique F\u00e9d\u00e9rale de Lausanne (EPFL), Information, Learning and Physics lab., CH-1015 Lausanne, Switzerland"
],
"familyname": "Krzakala"
},
{
"givennames": "Lenka",
"affiliations": [
"\u00c9cole Polytechnique F\u00e9d\u00e9rale de Lausanne (EPFL), Statistical Physics of Computation lab., CH-1015 Lausanne, Switzerland"
],
"familyname": "Zdeborova"
}
],
"title": "Expectation consistency for calibration of neural networks (code)",
"_oai": {
"id": "oai:materialscloud.org:1892"
},
"keywords": [
"MARVEL/P2",
"neural networks",
"uncertainty quantification",
"temperature scaling"
],
"publication_date": "Sep 19, 2023, 16:35:05",
"_files": [
{
"key": "expectation-consistency-master.zip",
"description": "Compressed files of the Github repository https://github.com/SPOC-group/expectation-consistency",
"checksum": "md5:fbcc77f1b6d304e8b48982e4f3bf8d55",
"size": 31243819
}
],
"references": [
{
"doi": "https://doi.org/10.48550/arXiv.2303.02644",
"citation": "L. Clarte, B. Loureiro, F. Krzakala, L. Zdeborova, PMLR 216, 443-453 (2023)",
"url": "https://arxiv.org/abs/2303.02644",
"type": "Journal reference"
}
],
"description": "Despite their incredible performance, it is well reported that deep neural networks tend to be overoptimistic about their prediction confidence. Finding effective and efficient calibration methods for neural networks is therefore an important endeavour towards better uncertainty quantification in deep learning. In this manuscript, we introduce a novel calibration technique named expectation consistency (EC), consisting of a post-training rescaling of the last layer weights by enforcing that the average validation confidence coincides with the average proportion of correct labels. First, we show that the EC method achieves similar calibration performance to temperature scaling (TS) across different neural network architectures and data sets, all while requiring similar validation samples and computational resources. However, we argue that EC provides a principled method grounded on a Bayesian optimality principle known as the Nishimori identity. Next, we provide an asymptotic characterization of both TS and EC in a synthetic setting and show that their performance crucially depends on the target function. In particular, we discuss examples where EC significantly outperforms TS.\nThis record provides the code for the paper \"Expectation consistency for calibration of neural networks\".",
"status": "published",
"license": "Creative Commons Attribution 4.0 International",
"conceptrecid": "1891",
"is_last": true,
"mcid": "2023.144",
"edited_by": 576,
"id": "1892",
"owner": 1130,
"license_addendum": null,
"doi": "10.24435/materialscloud:ws-p3"
},
"revision": 6,
"created": "2023-09-07T16:47:31.506265+00:00"
}