Conditional Alignment and Uniformity for Contrastive Learning with Continuous Proxy Labels - Equipe Image, Modélisation, Analyse, GEométrie, Synthèse Access content directly
Conference Papers Year : 2021

Conditional Alignment and Uniformity for Contrastive Learning with Continuous Proxy Labels

Abstract

Contrastive Learning has shown impressive results on natural and medical images, without requiring annotated data. However, a particularity of medical images is the availability of meta-data (such as age or sex) that can be exploited for learning representations. Here, we show that the recently proposed contrastive y-Aware InfoNCE loss, that integrates multi-dimensional meta-data, asymptotically optimizes two properties: conditional alignment and global uniformity. Similarly to [33], conditional alignment means that similar samples should have similar features, but conditionally on the meta-data. Instead, global uniformity means that the (normalized) features should be uniformly distributed on the unit hypersphere, independently of the meta-data. Here, we propose to define conditional uniformity, relying on the meta-data, that repel only samples with dissimilar metadata. We show that direct optimization of both conditional alignment and uniformity improves the representations, in terms of linear evaluation, on both CIFAR-100 and a brain MRI dataset.
Fichier principal
Vignette du fichier
2111.05643.pdf (545.66 Ko) Télécharger le fichier
Origin : Files produced by the author(s)

Dates and versions

hal-03523114 , version 1 (12-01-2022)

Identifiers

  • HAL Id : hal-03523114 , version 1

Cite

Benoit Dufumier, Pietro Gori, Julie Victor, Antoine Grigis, Edouard Duchesnay. Conditional Alignment and Uniformity for Contrastive Learning with Continuous Proxy Labels. Med-NeurIPS - Workshop NeurIPS, Dec 2021, Vancouver, Canada. ⟨hal-03523114⟩
155 View
74 Download

Share

Gmail Facebook X LinkedIn More