Augmentations vs Algorithms: What Works in Self-Supervised Learning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Morningstar, Warren, Bijamov, Alex, Duvarney, Chris, Friedman, Luke, Kalibhat, Neha, Liu, Luyang, Mansfield, Philip, Rojas-Gomez, Renan, Singhal, Karan, Green, Bradley, Prakash, Sushant
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911792434774016
author Morningstar, Warren
Bijamov, Alex
Duvarney, Chris
Friedman, Luke
Kalibhat, Neha
Liu, Luyang
Mansfield, Philip
Rojas-Gomez, Renan
Singhal, Karan
Green, Bradley
Prakash, Sushant
author_facet Morningstar, Warren
Bijamov, Alex
Duvarney, Chris
Friedman, Luke
Kalibhat, Neha
Liu, Luyang
Mansfield, Philip
Rojas-Gomez, Renan
Singhal, Karan
Green, Bradley
Prakash, Sushant
contents We study the relative effects of data augmentations, pretraining algorithms, and model architectures in Self-Supervised Learning (SSL). While the recent literature in this space leaves the impression that the pretraining algorithm is of critical importance to performance, understanding its effect is complicated by the difficulty in making objective and direct comparisons between methods. We propose a new framework which unifies many seemingly disparate SSL methods into a single shared template. Using this framework, we identify aspects in which methods differ and observe that in addition to changing the pretraining algorithm, many works also use new data augmentations or more powerful model architectures. We compare several popular SSL methods using our framework and find that many algorithmic additions, such as prediction networks or new losses, have a minor impact on downstream task performance (often less than $1\%$), while enhanced augmentation techniques offer more significant performance improvements ($2-4\%$). Our findings challenge the premise that SSL is being driven primarily by algorithmic improvements, and suggest instead a bitter lesson for SSL: that augmentation diversity and data / model scale are more critical contributors to recent advances in self-supervised learning.
format Preprint
id arxiv_https___arxiv_org_abs_2403_05726
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Augmentations vs Algorithms: What Works in Self-Supervised Learning
Morningstar, Warren
Bijamov, Alex
Duvarney, Chris
Friedman, Luke
Kalibhat, Neha
Liu, Luyang
Mansfield, Philip
Rojas-Gomez, Renan
Singhal, Karan
Green, Bradley
Prakash, Sushant
Machine Learning
Computer Vision and Pattern Recognition
We study the relative effects of data augmentations, pretraining algorithms, and model architectures in Self-Supervised Learning (SSL). While the recent literature in this space leaves the impression that the pretraining algorithm is of critical importance to performance, understanding its effect is complicated by the difficulty in making objective and direct comparisons between methods. We propose a new framework which unifies many seemingly disparate SSL methods into a single shared template. Using this framework, we identify aspects in which methods differ and observe that in addition to changing the pretraining algorithm, many works also use new data augmentations or more powerful model architectures. We compare several popular SSL methods using our framework and find that many algorithmic additions, such as prediction networks or new losses, have a minor impact on downstream task performance (often less than $1\%$), while enhanced augmentation techniques offer more significant performance improvements ($2-4\%$). Our findings challenge the premise that SSL is being driven primarily by algorithmic improvements, and suggest instead a bitter lesson for SSL: that augmentation diversity and data / model scale are more critical contributors to recent advances in self-supervised learning.
title Augmentations vs Algorithms: What Works in Self-Supervised Learning
topic Machine Learning
Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2403.05726