Preventing Dimensional Collapse in Contrastive Local Learning with Subsampling - Archive ouverte HAL Access content directly
Conference Papers Year : 2023

Preventing Dimensional Collapse in Contrastive Local Learning with Subsampling

Abstract

This paper presents an investigation of the challenges of training Deep Neural Networks (DNNs) via self-supervised objectives, using local learning as a parallelizable alternative to traditional backpropagation. In our approach, DNN are segmented into distinct blocks, each updated independently via gradients provided by small local auxiliary Neural Networks (NNs). Despite the evident computational benefits, extensive splits often result in performance degradation. Through analysis of a synthetic example, we identify a layer-wise dimensional collapse as a major factor behind such performance losses. To counter this, we propose a novel and straightforward sampling strategy based on blockwise feature-similarity, explicitly designed to evade such dimensional collapse.
Fichier principal
Vignette du fichier
main.pdf (1.32 Mo) Télécharger le fichier
Origin Files produced by the author(s)

Dates and versions

hal-04156218 , version 1 (20-07-2023)

Identifiers

  • HAL Id : hal-04156218 , version 1

Cite

Louis Fournier, Adeetya Patel, Michael Eickenberg, Edouard Oyallon, Eugene Belilovsky. Preventing Dimensional Collapse in Contrastive Local Learning with Subsampling. ICML 2023 Workshop on Localized Learning (LLW), Jul 2023, Honolulu (Hawaii), USA, United States. ⟨hal-04156218⟩
104 View
164 Download

Share

Gmail Mastodon Facebook X LinkedIn More