Timezone: »
In this paper, we study contrastive learning from an optimization perspective, aiming to analyze and address a fundamental issue of existing contrastive learning methods that either rely on a large batch size or a large dictionary of feature vectors. We consider a global objective for contrastive learning, which contrasts each positive pair with all negative pairs for an anchor point. From the optimization perspective, we explain why existing methods such as SimCLR require a large batch size in order to achieve a satisfactory result. In order to remove such requirement, we propose a memory-efficient Stochastic Optimization algorithm for solving the Global objective of Contrastive Learning of Representations, named SogCLR. We show that its optimization error is negligible under a reasonable condition after a sufficient number of iterations or is diminishing for a slightly different global contrastive objective. Empirically, we demonstrate that SogCLR with small batch size (e.g., 256) can achieve similar performance as SimCLR with large batch size (e.g., 8192) on self-supervised learning task on ImageNet-1K. We also attempt to show that the proposed optimization technique is generic and can be applied to solving other contrastive losses, e.g., two-way contrastive losses for bimodal contrastive learning. The proposed method is implemented in our open-sourced library LibAUC (www.libauc.org).
Author Information
Zhuoning Yuan (The University of Iowa)
Ph.D. student in Machine Learning
Yuexin Wu (Google)
Zi-Hao Qiu (Nanjing University)
Xianzhi Du (Google Inc.)
Lijun Zhang (Nanjing University)
Denny Zhou (Google Brain)
Tianbao Yang (The University of Iowa)
Related Events (a corresponding poster, oral, or spotlight)
-
2022 Poster: Provable Stochastic Optimization for Global Contrastive Learning: Small Batch Does Not Harm Performance »
Thu. Jul 21st through Fri the 22nd Room Hall E #329
More from the Same Authors
-
2023 Poster: Provable Multi-instance Deep AUC Maximization with Stochastic Pooling »
Dixian Zhu · Bokun Wang · Zhi Chen · Yaxing Wang · Milan Sonka · Xiaodong Wu · Tianbao Yang -
2023 Poster: Label Distributionally Robust Losses for Multi-class Classification: Consistency, Robustness and Adaptivity »
Dixian Zhu · Yiming Ying · Tianbao Yang -
2023 Poster: Generalization Analysis for Contrastive Representation Learning »
Yunwen Lei · Tianbao Yang · Yiming Ying · Ding-Xuan Zhou -
2023 Poster: Not All Semantics are Created Equal: Contrastive Self-supervised Learning with Automatic Temperature Individualization »
Zi-Hao Qiu · Quanqi Hu · Zhuoning Yuan · Denny Zhou · Lijun Zhang · Tianbao Yang -
2023 Poster: Learning Unnormalized Statistical Models via Compositional Optimization »
Wei Jiang · Jiayu Qin · Lingyu Wu · Changyou Chen · Tianbao Yang · Lijun Zhang -
2023 Poster: The Flan Collection: Designing Data and Methods for Effective Instruction Tuning »
Shayne Longpre · Le Hou · Tu Vu · Albert Webson · Hyung Won Chung · Yi Tay · Denny Zhou · Quoc Le · Barret Zoph · Jason Wei · Adam Roberts -
2023 Poster: Large Language Models Can Be Easily Distracted by Irrelevant Context »
Haoyue Shi · Xinyun Chen · Kanishka Misra · Nathan Scales · David Dohan · Ed Chi · Nathanael Schärli · Denny Zhou -
2023 Poster: Blockwise Stochastic Variance-Reduced Methods with Parallel Speedup for Multi-Block Bilevel Optimization »
Quanqi Hu · Zi-Hao Qiu · Zhishuai Guo · Lijun Zhang · Tianbao Yang -
2023 Poster: FeDXL: Provable Federated Learning for Deep X-Risk Optimization »
Zhishuai Guo · Rong Jin · Jiebo Luo · Tianbao Yang -
2022 Poster: A Simple yet Universal Strategy for Online Convex Optimization »
Lijun Zhang · Guanghui Wang · Jinfeng Yi · Tianbao Yang -
2022 Oral: A Simple yet Universal Strategy for Online Convex Optimization »
Lijun Zhang · Guanghui Wang · Jinfeng Yi · Tianbao Yang -
2022 Poster: GraphFM: Improving Large-Scale GNN Training via Feature Momentum »
Haiyang Yu · Limei Wang · Bokun Wang · Meng Liu · Tianbao Yang · Shuiwang Ji -
2022 Poster: Optimal Algorithms for Stochastic Multi-Level Compositional Optimization »
Wei Jiang · Bokun Wang · Yibo Wang · Lijun Zhang · Tianbao Yang -
2022 Poster: Large-scale Stochastic Optimization of NDCG Surrogates for Deep Learning with Provable Convergence »
Zi-Hao Qiu · Quanqi Hu · Yongjian Zhong · Lijun Zhang · Tianbao Yang -
2022 Poster: Finite-Sum Coupled Compositional Stochastic Optimization: Theory and Applications »
Bokun Wang · Tianbao Yang -
2022 Spotlight: GraphFM: Improving Large-Scale GNN Training via Feature Momentum »
Haiyang Yu · Limei Wang · Bokun Wang · Meng Liu · Tianbao Yang · Shuiwang Ji -
2022 Spotlight: Large-scale Stochastic Optimization of NDCG Surrogates for Deep Learning with Provable Convergence »
Zi-Hao Qiu · Quanqi Hu · Yongjian Zhong · Lijun Zhang · Tianbao Yang -
2022 Spotlight: Finite-Sum Coupled Compositional Stochastic Optimization: Theory and Applications »
Bokun Wang · Tianbao Yang -
2022 Spotlight: Optimal Algorithms for Stochastic Multi-Level Compositional Optimization »
Wei Jiang · Bokun Wang · Yibo Wang · Lijun Zhang · Tianbao Yang -
2022 Poster: When AUC meets DRO: Optimizing Partial AUC for Deep Learning with Non-Convex Convergence Guarantee »
Dixian Zhu · Gang Li · Bokun Wang · Xiaodong Wu · Tianbao Yang -
2022 Spotlight: When AUC meets DRO: Optimizing Partial AUC for Deep Learning with Non-Convex Convergence Guarantee »
Dixian Zhu · Gang Li · Bokun Wang · Xiaodong Wu · Tianbao Yang -
2021 Poster: SpreadsheetCoder: Formula Prediction from Semi-structured Context »
Xinyun Chen · Petros Maniatis · Rishabh Singh · Charles Sutton · Hanjun Dai · Max Lin · Denny Zhou -
2021 Poster: LEGO: Latent Execution-Guided Reasoning for Multi-Hop Question Answering on Knowledge Graphs »
Hongyu Ren · Hanjun Dai · Bo Dai · Xinyun Chen · Michihiro Yasunaga · Haitian Sun · Dale Schuurmans · Jure Leskovec · Denny Zhou -
2021 Poster: Stability and Generalization of Stochastic Gradient Methods for Minimax Problems »
Yunwen Lei · Zhenhuan Yang · Tianbao Yang · Yiming Ying -
2021 Spotlight: SpreadsheetCoder: Formula Prediction from Semi-structured Context »
Xinyun Chen · Petros Maniatis · Rishabh Singh · Charles Sutton · Hanjun Dai · Max Lin · Denny Zhou -
2021 Spotlight: LEGO: Latent Execution-Guided Reasoning for Multi-Hop Question Answering on Knowledge Graphs »
Hongyu Ren · Hanjun Dai · Bo Dai · Xinyun Chen · Michihiro Yasunaga · Haitian Sun · Dale Schuurmans · Jure Leskovec · Denny Zhou -
2021 Oral: Stability and Generalization of Stochastic Gradient Methods for Minimax Problems »
Yunwen Lei · Zhenhuan Yang · Tianbao Yang · Yiming Ying -
2021 Poster: Federated Deep AUC Maximization for Hetergeneous Data with a Constant Communication Complexity »
Zhuoning Yuan · Zhishuai Guo · Yi Xu · Yiming Ying · Tianbao Yang -
2021 Spotlight: Federated Deep AUC Maximization for Hetergeneous Data with a Constant Communication Complexity »
Zhuoning Yuan · Zhishuai Guo · Yi Xu · Yiming Ying · Tianbao Yang -
2020 Poster: Good Subnetworks Provably Exist: Pruning via Greedy Forward Selection »
Mao Ye · Chengyue Gong · Lizhen Nie · Denny Zhou · Adam Klivans · Qiang Liu -
2020 Poster: Communication-Efficient Distributed Stochastic AUC Maximization with Deep Neural Networks »
Zhishuai Guo · Mingrui Liu · Zhuoning Yuan · Li Shen · Wei Liu · Tianbao Yang -
2020 Poster: Go Wide, Then Narrow: Efficient Training of Deep Thin Networks »
Denny Zhou · Mao Ye · Chen Chen · Tianjian Meng · Mingxing Tan · Xiaodan Song · Quoc Le · Qiang Liu · Dale Schuurmans -
2020 Poster: Quadratically Regularized Subgradient Methods for Weakly Convex Optimization with Weakly Convex Constraints »
Runchao Ma · Qihang Lin · Tianbao Yang -
2020 Poster: Stochastic Optimization for Non-convex Inf-Projection Problems »
Yan Yan · Yi Xu · Lijun Zhang · Wang Xiaoyu · Tianbao Yang -
2019 Poster: Stochastic Optimization for DC Functions and Non-smooth Non-convex Regularizers with Non-asymptotic Convergence »
Yi Xu · Qi Qi · Qihang Lin · rong jin · Tianbao Yang -
2019 Oral: Stochastic Optimization for DC Functions and Non-smooth Non-convex Regularizers with Non-asymptotic Convergence »
Yi Xu · Qi Qi · Qihang Lin · rong jin · Tianbao Yang -
2019 Poster: Katalyst: Boosting Convex Katayusha for Non-Convex Problems with a Large Condition Number »
Zaiyi Chen · Yi Xu · Haoyuan Hu · Tianbao Yang -
2019 Oral: Katalyst: Boosting Convex Katayusha for Non-Convex Problems with a Large Condition Number »
Zaiyi Chen · Yi Xu · Haoyuan Hu · Tianbao Yang -
2018 Poster: Dynamic Regret of Strongly Adaptive Methods »
Lijun Zhang · Tianbao Yang · rong jin · Zhi-Hua Zhou -
2018 Poster: SADAGRAD: Strongly Adaptive Stochastic Gradient Methods »
Zaiyi Chen · Yi Xu · Enhong Chen · Tianbao Yang -
2018 Poster: Level-Set Methods for Finite-Sum Constrained Convex Optimization »
Qihang Lin · Runchao Ma · Tianbao Yang -
2018 Oral: Level-Set Methods for Finite-Sum Constrained Convex Optimization »
Qihang Lin · Runchao Ma · Tianbao Yang -
2018 Oral: SADAGRAD: Strongly Adaptive Stochastic Gradient Methods »
Zaiyi Chen · Yi Xu · Enhong Chen · Tianbao Yang -
2018 Oral: Dynamic Regret of Strongly Adaptive Methods »
Lijun Zhang · Tianbao Yang · rong jin · Zhi-Hua Zhou -
2018 Poster: Fast Stochastic AUC Maximization with $O(1/n)$-Convergence Rate »
Mingrui Liu · Xiaoxuan Zhang · Zaiyi Chen · Xiaoyu Wang · Tianbao Yang -
2018 Oral: Fast Stochastic AUC Maximization with $O(1/n)$-Convergence Rate »
Mingrui Liu · Xiaoxuan Zhang · Zaiyi Chen · Xiaoyu Wang · Tianbao Yang -
2017 Poster: Stochastic Convex Optimization: Faster Local Growth Implies Faster Global Convergence »
Yi Xu · Qihang Lin · Tianbao Yang -
2017 Poster: A Richer Theory of Convex Constrained Optimization with Reduced Projections and Improved Rates »
Tianbao Yang · Qihang Lin · Lijun Zhang -
2017 Talk: A Richer Theory of Convex Constrained Optimization with Reduced Projections and Improved Rates »
Tianbao Yang · Qihang Lin · Lijun Zhang -
2017 Talk: Stochastic Convex Optimization: Faster Local Growth Implies Faster Global Convergence »
Yi Xu · Qihang Lin · Tianbao Yang