Timezone: »
Large pre-trained models such as CLIP offer consistent accuracy across a range of data distributions when performing zero-shot inference (i.e., without fine-tuning on a specific dataset). Although existing fine-tuning approaches substantially improve accuracy in-distribution, they also reduce out-of-distribution robustness. We address this tension by introducing a simple and effective method for improving robustness: ensembling the weights of the zero-shot and fine-tuned models (WiSE-FT). Compared to standard fine-tuning, WiSE-FT provides large accuracy improvements out-of-distribution, while matching or improving in-distribution accuracy. On ImageNet (in-distribution) and five derived distribution shifts, WiSE-FT improves out-of-distribution accuracy by 2 to 10 percentage points (pp) while increasing in-distribution accuracy by nearly 1 pp relative to standard fine-tuning. WiSE-FT achieves similarly large robustness improvements (2 to 15 pp) on a diverse set of six further distribution shifts, and in-distribution accuracy gains of 0.8 to 3.3 pp compared to standard fine-tuning on seven commonly used transfer learning datasets. These improvements come at no additional computational cost during fine-tuning or inference.
Author Information
Mitchell Wortsman (University of Washington, Allen Institute for Artificial Intelligence)
Gabriel Ilharco (Department of Computer Science, University of Washington)
Jong Wook Kim (OpenAI)
Jong Wook Kim is a member of technical staff at OpenAI, where he worked on GPT-2 output detection, Jukebox, and CLIP. His research interests include representation learning and generative modeling of audio and music, as well as its applications to multimodal deep learning. Prior to OpenAI, he completed a Ph.D. in music technology from NYU, which focused on automatic music transcription. He also worked as a research scientist intern at Pandora and Spotify, and as a software engineer at Kakao and NCSOFT.
Mike Li (Columbia University)
Hanna Hajishirzi (University of Washington)
Ali Farhadi (University of Washington, Allen Institute for Artificial Intelligence)
Hongseok Namkoong (Stanford University)
Ludwig Schmidt (University of Washington)
More from the Same Authors
-
2021 : Are We Learning Yet? A Meta Review of Evaluation Failures Across Machine Learning »
Thomas Liao · Rohan Taori · Deborah Raji · Ludwig Schmidt -
2021 : Do ImageNet Classifiers Generalize to ImageNet? »
Benjamin Recht · Becca Roelofs · Ludwig Schmidt · Vaishaal Shankar -
2021 : Evaluating Machine Accuracy on ImageNet »
Vaishaal Shankar · Becca Roelofs · Horia Mania · Benjamin Recht · Ludwig Schmidt -
2021 : Measuring Robustness to Natural Distribution Shifts in Image Classification »
Rohan Taori · Achal Dave · Vaishaal Shankar · Nicholas Carlini · Benjamin Recht · Ludwig Schmidt -
2023 Poster: Stable and low-precision training for large-scale vision-language models »
Mitchell Wortsman · Tim Dettmers · Luke Zettlemoyer · Ari Morcos · Ali Farhadi · Ludwig Schmidt -
2023 Poster: AdANNS: A Framework for Adaptive Semantic Search »
Aniket Rege · Aditya Kusupati · Sharan Ranjit S · Alan Fan · Qingqing Cao · Sham Kakade · Prateek Jain · Ali Farhadi -
2023 Poster: Localized Symbolic Knowledge Distillation for Visual Commonsense Models »
Jae Sung Park · Jack Hessel · Khyathi Chandu · Paul Pu Liang · Ximing Lu · Qiuyuan Huang · Peter West · Jianfeng Gao · Ali Farhadi · Yejin Choi -
2023 Poster: Neural Priming for Sample-Efficient Adaptation »
Matthew Wallingford · Vivek Ramanujan · Alex Fang · Aditya Kusupati · Roozbeh Mottaghi · Aniruddha Kembhavi · Ludwig Schmidt · Ali Farhadi -
2023 Poster: On the Connection between Pre-training Data Diversity and Fine-tuning Robustness »
Vivek Ramanujan · Thao Nguyen · Sewoong Oh · Ali Farhadi · Ludwig Schmidt -
2023 Poster: DataComp: In search of the next generation of multimodal datasets »
Samir Yitzhak Gadre · Gabriel Ilharco · Alex Fang · Jonathan Hayase · Georgios Smyrnis · Thao Nguyen · Ryan Marten · Mitchell Wortsman · Dhruba Ghosh · Jieyu Zhang · Eyal Orgad · Rahim Entezari · Giannis Daras · Sarah Pratt · Vivek Ramanujan · Yonatan Bitton · Kalyani Marathe · Stephen Mussmann · Richard Vencu · Mehdi Cherti · Ranjay Krishna · Pang Wei Koh · Olga Saukh · Alexander Ratner · Shuran Song · Hannaneh Hajishirzi · Ali Farhadi · Romain Beaumont · Sewoong Oh · Alex Dimakis · Jenia Jitsev · Yair Carmon · Vaishaal Shankar · Ludwig Schmidt -
2023 Poster: Improving multimodal datasets with image captioning »
Thao Nguyen · Samir Yitzhak Gadre · Gabriel Ilharco · Sewoong Oh · Ludwig Schmidt -
2023 Poster: Objaverse-XL: A Colossal Universe of 3D Objects »
Matt Deitke · Ruoshi Liu · Matthew Wallingford · Huong Ngo · Oscar Michel · Aditya Kusupati · Alan Fan · Christian Laforte · Vikram Voleti · Samir Yitzhak Gadre · Eli VanderBilt · Aniruddha Kembhavi · Carl Vondrick · Georgia Gkioxari · Kiana Ehsani · Ludwig Schmidt · Ali Farhadi -
2023 Oral: DataComp: In search of the next generation of multimodal datasets »
Samir Yitzhak Gadre · Gabriel Ilharco · Alex Fang · Jonathan Hayase · Georgios Smyrnis · Thao Nguyen · Ryan Marten · Mitchell Wortsman · Dhruba Ghosh · Jieyu Zhang · Eyal Orgad · Rahim Entezari · Giannis Daras · Sarah Pratt · Vivek Ramanujan · Yonatan Bitton · Kalyani Marathe · Stephen Mussmann · Richard Vencu · Mehdi Cherti · Ranjay Krishna · Pang Wei Koh · Olga Saukh · Alexander Ratner · Shuran Song · Hannaneh Hajishirzi · Ali Farhadi · Romain Beaumont · Sewoong Oh · Alex Dimakis · Jenia Jitsev · Yair Carmon · Vaishaal Shankar · Ludwig Schmidt -
2022 Poster: Patching open-vocabulary models by interpolating weights »
Gabriel Ilharco · Mitchell Wortsman · Samir Yitzhak Gadre · Shuran Song · Hannaneh Hajishirzi · Simon Kornblith · Ali Farhadi · Ludwig Schmidt -
2022 Poster: LAION-5B: An open large-scale dataset for training next generation image-text models »
Christoph Schuhmann · Romain Beaumont · Richard Vencu · Cade Gordon · Ross Wightman · Mehdi Cherti · Theo Coombes · Aarush Katta · Clayton Mullis · Mitchell Wortsman · Patrick Schramowski · Srivatsa Kundurthy · Katherine Crowson · Ludwig Schmidt · Robert Kaczmarczyk · Jenia Jitsev -
2022 Poster: Quality Not Quantity: On the Interaction between Dataset Design and Robustness of CLIP »
Thao Nguyen · Gabriel Ilharco · Mitchell Wortsman · Sewoong Oh · Ludwig Schmidt -
2022 Poster: Matryoshka Representation Learning »
Aditya Kusupati · Gantavya Bhatt · Aniket Rege · Matthew Wallingford · Aditya Sinha · Vivek Ramanujan · William Howard-Snyder · Kaifeng Chen · Sham Kakade · Prateek Jain · Ali Farhadi -
2021 Oral: Retiring Adult: New Datasets for Fair Machine Learning »
Frances Ding · Moritz Hardt · John Miller · Ludwig Schmidt -
2021 Oral: MERLOT: Multimodal Neural Script Knowledge Models »
Rowan Zellers · Ximing Lu · Jack Hessel · Youngjae Yu · Jae Sung Park · Jize Cao · Ali Farhadi · Yejin Choi -
2021 Poster: MERLOT: Multimodal Neural Script Knowledge Models »
Rowan Zellers · Ximing Lu · Jack Hessel · Youngjae Yu · Jae Sung Park · Jize Cao · Ali Farhadi · Yejin Choi -
2021 Poster: Retiring Adult: New Datasets for Fair Machine Learning »
Frances Ding · Moritz Hardt · John Miller · Ludwig Schmidt -
2021 Poster: Characterizing Generalization under Out-Of-Distribution Shifts in Deep Metric Learning »
Timo Milbich · Karsten Roth · Samarth Sinha · Ludwig Schmidt · Marzyeh Ghassemi · Bjorn Ommer -
2021 Poster: LLC: Accurate, Multi-purpose Learnt Low-dimensional Binary Codes »
Aditya Kusupati · Matthew Wallingford · Vivek Ramanujan · Raghav Somani · Jae Sung Park · Krishna Pillutla · Prateek Jain · Sham Kakade · Ali Farhadi -
2021 Poster: Evaluating model performance under worst-case subpopulations »
Mike Li · Hongseok Namkoong · Shangzhou Xia -
2021 : Techniques and Conclusion »
Lilian Weng · Jong Wook Kim -
2021 : Pretext tasks »
Jong Wook Kim -
2021 : Pretext tasks (vision) »
Jong Wook Kim -
2021 : Early Work »
Jong Wook Kim -
2021 Tutorial: Self-Supervised Learning: Self-Prediction and Contrastive Learning »
Lilian Weng · Jong Wook Kim -
2020 Poster: Supermasks in Superposition »
Mitchell Wortsman · Vivek Ramanujan · Rosanne Liu · Aniruddha Kembhavi · Mohammad Rastegari · Jason Yosinski · Ali Farhadi -
2019 : Coffee break, posters, and 1-on-1 discussions »
Yangyi Lu · Daniel Chen · Hongseok Namkoong · Marie Charpignon · Maja Rudolph · Amanda Coston · Julius von Kügelgen · Niranjani Prasad · Paramveer Dhillon · Yunzong Xu · Yixin Wang · Alexander Markham · David Rohde · Rahul Singh · Zichen Zhang · Negar Hassanpour · Ankit Sharma · Ciarán Lee · Jean Pouget-Abadie · Jesse Krijthe · Divyat Mahajan · Nan Rosemary Ke · Peter Wirnsberger · Vira Semenova · Dmytro Mykhaylov · Dennis Shen · Kenta Takatsu · Liyang Sun · Jeremy Yang · Alexander Franks · Pak Kan Wong · Tauhid Zaman · Shira Mitchell · min kyoung kang · Qi Yang -
2019 : Poster Spotlights »
Hongseok Namkoong · Marie Charpignon · Maja Rudolph · Amanda Coston · Yuta Saito · Paramveer Dhillon · Alexander Markham -
2019 Poster: Defending Against Neural Fake News »
Rowan Zellers · Ari Holtzman · Hannah Rashkin · Yonatan Bisk · Ali Farhadi · Franziska Roesner · Yejin Choi -
2019 Poster: Model Similarity Mitigates Test Set Overuse »
Horia Mania · John Miller · Ludwig Schmidt · Moritz Hardt · Benjamin Recht -
2019 Poster: Unlabeled Data Improves Adversarial Robustness »
Yair Carmon · Aditi Raghunathan · Ludwig Schmidt · John Duchi · Percy Liang -
2019 Poster: A Meta-Analysis of Overfitting in Machine Learning »
Becca Roelofs · Vaishaal Shankar · Benjamin Recht · Sara Fridovich-Keil · Moritz Hardt · John Miller · Ludwig Schmidt -
2019 Poster: Discovering Neural Wirings »
Mitchell Wortsman · Ali Farhadi · Mohammad Rastegari -
2017 Poster: Variance-based Regularization with Convex Objectives »
Hongseok Namkoong · John Duchi -
2017 Oral: Variance-based Regularization with Convex Objectives »
Hongseok Namkoong · John Duchi