CoChat
Sign Up
Cited by 2,484Advanced Neural Network ApplicationsDomain Adaptation and Few-Shot LearningMachine Learning and Data Classification

Revisiting Unreasonable Effectiveness of Data in Deep Learning Era

Open access
Sun, Chen, Shrivastava, Abhinav, Singh, Saurabh, Gupta, Abhinav·2017·2017 IEEE International Conference on Computer Vision (ICCV)
DOI 10.1109/iccv.2017.97
Open this paper in CoChat — free accountOpen PDF

Abstract

The success of deep learning in vision can be attributed to: (a) models with high capacity; (b) increased computational power; and (c) availability of large-scale labeled data. Since 2012, there have been significant advances in representation capabilities of the models and computational capabilities of GPUs. But the size of the biggest dataset has surprisingly remained constant. What will happen if we increase the dataset size by 10 × or 100 × ? This paper takes a step towards clearing the clouds of mystery surrounding the relationship between ‘enormous data’ and visual deep learning. By exploiting the JFT-300M dataset which has more than 375M noisy labels for 300M images, we investigate how the performance of current vision tasks would change if this data was used for representation learning. Our paper delivers some surprising (and some expected) findings. First, we find that the performance on vision tasks increases logarithmically based on volume of training data size. Second, we show that representation learning (or pre-training) still holds a lot of promise. One can improve performance on many vision tasks by just training a better base model. Finally, as expected, we present new state-of-the-art results for different vision tasks including image classification, object detection, semantic segmentation and human pose estimation. Our sincere hope is that this inspires vision community to not undervalue the data and develop collective efforts in building larger datasets.

Integrity signals

Registry

Crossref · 10.1109/iccv.2017.97

Cites

11 of 11 references verified0 citation errors · 0 not checkable

Statistics

Recomputing p-values needs the paper’s full text.

Claims the paper makes

We couldn’t retrieve this paper’s full text, so its claims and self-check weren’t run. The registry and citation checks still apply.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • ViTs outperform CNNs on large datasets, such as ImageNet and JFT.

    “As a result, ViTs outperform CNNs on large datasets, such as ImageNet and JFT (Sun et al., 2017).”

    From How Do Vision Transformers Work? (2022) · cites this paper as Sun et al., 2017

  • Modern deep learning models feature millions of parameters that need to be learned from sufficiently large curated data sets in order to achieve clinical-grade accuracy, while being safe, fair, equitable and generalising well to unseen data.

    “Modern DL models feature millions of parameters that need to be learned from sufficiently large curated data sets in order to achieve clinical-grade accuracy, while being safe, fair, equitable and generalising well to unseen data2–5.”

    From The future of digital health with federated learning (2020) · cites this paper as [5]

  • ViT models achieve better performance than state-of-the-art ResNets of similar capacity when both are pre-trained on sufficiently large datasets such as JFT-300M.

    “In this paper, we focus on one particular Transformer architecture-the Visual Transformer (ViT) introduced by Dosovitskiy et al. [9]-because it was shown to achieve better performance than state-of-the-art residual networks (ResNets) [17] of similar capacity, when both are pre-trained on sufficiently large datasets [28], such as JFT-300M [43].”

    From Understanding Robustness of Transformers for Image Classification (2021) · cites this paper as [43]

  • ViT-B/32 shows limited improvements when pretraining on JFT-300M vs ImageNet.

    “the smallest model (ViT-B/32) shows limited improvements when pretraining on JFT-300M [40] vs. the ImageNet Large Scale Visual Recognition Challenge 2012 dataset [12, 37]”

    From Do Vision Transformers See Like Convolutional Neural Networks? (2021) · cites this paper as [40]

  • ViT-H/14 achieves state of the art when pretrained on JFT-300M.

    “while the largest, ViT-H/14, achieves state of the art when pretrained on JFT-300M [40].”

    From Do Vision Transformers See Like Convolutional Neural Networks? (2021) · cites this paper as [40]

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Authors
Sun, C., Shrivastava, A., Singh, S., & Gupta, A.
Affiliations
Google (United States) (US)
ORCID
Venue
2017 IEEE International Conference on Computer Vision (ICCV)
Publisher
IEEE
Published
2017-10
Pages
843-852
DOI
10.1109/iccv.2017.97
OpenAlex
W2962843773
Cited by
2,484
References
11
Citations by year
2026: 107 · 2025: 298 · 2024: 371 · 2023: 382 · 2022: 395 · 2021: 392 · 2020: 247 · 2019: 163 · 2018: 105 · 2017: 22 · 2016: 1
FWCI
43.03
Type
conference paper
Language
en
Open access
Locations
  • IEEE/CVF International Conference on Computer Vision (ICCV) (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning and Data Classification
Keywords
training data size, representation learning, pre-training, label noise, image classification, object detection, semantic segmentation, human pose estimation, large-scale image databases

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review