A NeurIPS 2021 workshop on
ImageNet: past, present, and future
Summary
Since its release in 2009, ImageNet has played an instrumental role in the development of deep learning architectures for computer vision, enabling neural networks to greatly outperform hand-crafted visual representations. ImageNet also quickly became the go-to benchmark for model architectures and training techniques which eventually reach far beyond image classification. Today’s models are getting close to “solving” the benchmark. Models trained on ImageNet have been used as strong initialization for numerous downstream tasks. The ImageNet dataset has even been used for tasks going way beyond its initial purpose of the training of classification models. It has been leveraged and reinvented for tasks such as few-shot learning, self-supervised learning and semi-supervised learning. Interesting re-creation of the ImageNet benchmark enables the evaluation of novel challenges like robustness, bias, or concept generalization. More accurate labels have been provided. About 10 years later, ImageNet symbolizes a decade of staggering advances in computer vision, deep learning, and artificial intelligence.
We believe now is a good time to discuss what’s next: Did we solve ImageNet? If not, what is there left to do? What are the main lessons learnt thanks to this benchmark? What should the next generation of ImageNet-like benchmarks encompass? Is language supervision a promising alternative? How can we reflect on the diverse requirements for good datasets and models, such as fairness, privacy, security, generalization, scale, and efficiency?
Schedule & contents
The workshop has taken place on Monday 13 December - Tuesday 14 December 2021, depending on your time zone.
The recordings are available below.
How to attend the workshop (closed)
You can attend the live event through the NeurIPS virtual site for our workshop. Access to the page requires registration at NeurIPS. You will find links to the Zoom and Gathertown sessions on the virtual site.
You can watch the Zoom events either through the livestream on the virtual page or on Zoom.
You can watch the Livestream events can only through the livestream on the virtual page.
You can join the poster session here on Gathertown.
How to actively participate in the workshop (closed)
We highly encourage the active participation of the attendees. Attendees can participate in various ways.
Invited talks, spotlight talks and panel sessions
Register questions and topics for discussion before the workshop: Google forms (link)
Post questions and topics for discussion during the live events: Rocketchat (requires registration)
Invited speakers will answer the selected questions during the panel sessions.
Poster sessions
Join the Gathertown during the poster sessions and interact with the authors, speakers, other attendees, and organisers.
After the workshop
We plan to publish the recorded events from the workshop.
The NeurIPS Code of Conduct applies to our workshop.
Important dates (closed)
Submission system open: 26 August 2021
Deadline for tracks #1 and #2: 18 September 2021 25 September 2021, 23:59 AOE (final; no extension)
Deadline for track #3: 5 December 2021 (Please submit to imagenetworkshoppc@gmail.com)
Author notification: 15 October 2021 22 October 2021
Workshop date: 13 December 2021
Feel free to join the mailing-list to be notified of updates, or simply check back later.
Call for papers (closed)
We call for research work on and around ImageNet to be presented at the workshop, see “Scope and topics” below for details. There are three submission tracks. All accepted papers, including the ones for the main track, are non-archival; that is, there are no formally published proceedings. Nonetheless, we will make all accepted papers available on OpenReview.
1. Main track
We welcome original research papers in the main track. The submissions must be at most 5 pages in length, excluding the references and appendix, and use our modified NeurIPS 2021 style file (see below). Reviewers are not obliged to read the appendix. Please submit a single PDF file that includes the main paper, references, and any appendix. The review process is double-blind, so please ensure that the submission is appropriately anonymised.
2. Extended abstract track
We call for extended abstract submissions that present open problems, novel-yet-not-thoroughly-tested ideas, and any perspective on the future of ImageNet and related research. The submissions must be at most 3 pages in length, excluding the references, and use our modified NeurIPS 2021 style file (see below). The review process is double-blind, so please ensure that the submission is appropriately anonymised.
3. Published work track
To encourage lively discussion around ImageNet, we allow the presentation of previously published papers, if they are highly relevant to the topic. Please note that it is not possible to submit papers accepted at the NeurIPS’21 main conference, as per the NeurIPS guidelines this year. There is no formal style and page limit on this track; authors may submit a published paper as-is. The review process is single-blind. Authors of the accepted papers will be able to present their work during the workshop.
Prizes for best papers
We will award the best papers for the main track (track 1) with a cash prize of 1000 USD. Stay tuned for further updates on the prizes!
Reviewer registration for authors
Given the expected size of the workshop, we will likely be short of reviewers. Authors will be given an option to register a subset of authors as reviewers in the OpenReview submission form. We strongly encourage at least one author for each paper to be registered as a reviewer. The paper reviewing will take place between late September and early October. We aim to assign at most two papers per reviewer.
Links and resources
Submission link: Closed. Track 3 is still open; please submit to imagenetworkshoppc@gmail.com.
Template files:
Scope and topics
All three tracks welcome any submission on the topics below. The list is non-exhaustive; if unsure, please contact the organizers at imagenet21@googlegroups.com.
ImageNet as a benchmark
How is ImageNet serving the community as a benchmark?
Did we solve the ImageNet classification task?
How can we spot and remedy the remaining errors?
What are the remaining challenges not captured by ImageNet?
How can we measure the research progress on robustness, bias, and other out-of-domain generalization issues using ImageNet?
How to fix errors in ImageNet labels and compare results across different labels?
ImageNet as a pre-training dataset
Is ImageNet still the best dataset for pre-training?
What could a better pre-training dataset look like? What are good properties?
Zero- and few-shot learning on or using ImageNet
Unsupervised and self-supervised learning
Semi-supervised learning
Transfer learning
Explorative and innovative ideas even without state-of-the-art (SOTA) results
We encourage submissions that are groundbreaking but are difficult to get published at conferences because they do not get SOTA performances.
Beyond ImageNet
Recipes for building a large-scale dataset
Socially responsible benchmarks and datasets: fairness, privacy, security, robustness, scalability, efficiency, and environmental friendliness
Efficiency and quality aspects of annotations and crowdsourcing
Going beyond classification labels for supervising visual models - e.g. language descriptions for images
Post-ImageNet benchmarks and datasets
Feel free to join the mailing-list to be notified of updates, or simply check back later.
Accepted papers
Best paper award (1000 USD sponsored by Naver)
ResNet strikes back: An improved training procedure in timm. Ross Wightman, Hugo Touvron, Herve Jegou.
Spotlight
ResNet strikes back: An improved training procedure in timm. Ross Wightman, Hugo Touvron, Herve Jegou.
Learning Background Invariance Improves Generalization and Robustness in Self Supervised Learning on ImageNet and Beyond. Chaitanya Ryali, David J. Schwab, Ari S. Morcos.
Posters
One Pass ImageNet. Huiyi Hu, Ang Li, Daniele Calandriello, Dilan Gorur.
ShapeY: Measuring Shape Recognition Capacity Using Nearest Neighbor Matching. Jong Woo Nam, Amanda Sofie Rios, Bartlett Mel.
How Well Does Self-Supervised Pre-Training Perform with Streaming ImageNet? Dapeng Hu, Shipeng Yan, Qizhengqiu Lu, Lanqing Hong, Hailin Hu, Yifan Zhang, Zhenguo Li, Jiashi Feng.
Transfer learning with fewer ImageNet classes. Michal Kucer, Diane Oyen.
ImageNet suffers from dichotomous data difficulty. Kristof Meding, Luca M. Schulze Buschoff, Robert Geirhos, Felix A. Wichmann.
Evaluating Adversarial Attacks on ImageNet: A Reality Check on Misclassification Classes. Utku Ozbulak, Maura Pintor, Arnout Van Messem, Wesley De Neve.
NAM: Normalization-based Attention Module. Yichao Liu, Zongru Shao, Yueyang Teng, Nico Hoffmann.
Published Work Track
Visually Grounded Reasoning across Languages and Cultures. Best Long Paper @ EMNLP 2021. Fangyu Liu, Emanuele Bugliarello, Edoardo Ponti, Siva Reddy, Nigel Collier, Desmond Elliott.
Evaluating Machine Accuracy on ImageNet. ICML 2020. Vaishaal Shankar, Rebecca Roelofs, Horia Mania, Alex Fang, Benjamin Recht, Ludwig Schmidt.
Measuring Robustness to Natural Distribution Shifts in Image Classification. Spotlight @ NeurIPS 2020. Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, Ludwig Schmidt.
Concept Generalization in Visual Representation Learning. ICCV 2021. Mert Bülent Sarıyıldız, Yannis Kalantidis, Diane Larlus, Karteek Alahari.
Do ImageNet Classifiers Generalize to ImageNet? ICML 2019. Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, Vaishaal Shankar.
IIRC: Incremental Implicitly-Refined Classification. CVPR 2021. Mohamed Abdelsalam, Mojtaba Faramarzi, Shagun Sodhani, Sarath Chandar.
PASS: An ImageNet replacement for self-supervised pretraining without humans. NeurIPS Track on Datasets and Benchmarks 2021. Yuki M Asano, Christian Rupprecht, Andrew Zisserman, Andrea Vedaldi.
An ecologically motivated image dataset for deep learning yields better models of human vision. PNAS 2021. Johannes Mehrer, Courtney J. Spoerer, Emer C. Jones, Nikolaus Kriegeskorte, and Tim C. Kietzmann.
Enriching ImageNet with Human Similarity Judgments and Psychological Embeddings. CVPR 2021. Brett D. Roads and Bradley C. Love.
Speakers
Matthias Bethge
Univ. of Tübingen
Emily Denton
Google Research
Vittorio Ferrari
Google Research
Alex Hanna
Google Research
Daniel
Hendrycks
UC Berkeley
Alex Kolesnikov
Google Brain
Sharon
Yixuan Li
UW Madison
Rebecca Roelofs
Google Brain
Olga Russakovsky
Princeton University
Shibani Santurkar
MIT
Dawn Song
UC Berkeley
Kaiyu Yang Princeton University
Ross Wightman
Independent Researcher
Organizers
You can contact the organizing team via imagenet21@googlegroups.com, and subscribe to the mailing-list to receive organizational updates.
Zeynep Akata
Univ. of Tübingen
Lucas Beyer
Google Brain
Sanghyuk Chun
Naver AI Lab
A. Sophia Koepke
Univ. of Tübingen
Diane Larlus
Naver Labs Europe
Seong Joon Oh
Naver AI Lab
Rafael Rezende
Naver Labs Europe
Sangdoo Yun
Naver AI Lab
Xiaohua Zhai
Google Brain