Collaboration of Experts: Achieving 80% Top-1 Accuracy on ImageNet with 100M FLOPs

Yikang Zhang, Zhuo Chen, Zhao Zhong
Proceedings of the 39th International Conference on Machine Learning, PMLR 162:26068-26084, 2022.

Abstract

In this paper, we propose a Collaboration of Experts (CoE) framework to assemble the expertise of multiple networks towards a common goal. Each expert is an individual network with expertise on a unique portion of the dataset, contributing to the collective capacity. Given a sample, delegator selects an expert and simultaneously outputs a rough prediction to trigger potential early termination. For each model in CoE, we propose a novel training algorithm with two major components: weight generation module (WGM) and label generation module (LGM). It fulfills the co-adaptation of experts and delegator. WGM partitions the training data into portions based on delegator via solving a balanced transportation problem, then impels each expert to focus on one portion by reweighting the losses. LGM generates the label to constitute the loss of delegator for expert selection. CoE achieves the state-of-the-art performance on ImageNet, 80.7% top-1 accuracy with 194M FLOPs. Combined with PWLU and CondConv, CoE further boosts the accuracy to 80.0% with only 100M FLOPs for the first time. Furthermore, experiment results on the translation task also demonstrate the strong generalizability of CoE. CoE is hardware-friendly, yielding a 3 6x acceleration compared with existing conditional computation approaches.

Cite this Paper

BibTeX
@InProceedings{pmlr-v162-zhang22c, title = {Collaboration of Experts: Achieving 80% Top-1 Accuracy on {I}mage{N}et with 100{M} {FLOP}s}, author = {Zhang, Yikang and Chen, Zhuo and Zhong, Zhao}, booktitle = {Proceedings of the 39th International Conference on Machine Learning}, pages = {26068--26084}, year = {2022}, editor = {Chaudhuri, Kamalika and Jegelka, Stefanie and Song, Le and Szepesvari, Csaba and Niu, Gang and Sabato, Sivan}, volume = {162}, series = {Proceedings of Machine Learning Research}, month = {17--23 Jul}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v162/zhang22c/zhang22c.pdf}, url = {https://proceedings.mlr.press/v162/zhang22c.html}, abstract = {In this paper, we propose a Collaboration of Experts (CoE) framework to assemble the expertise of multiple networks towards a common goal. Each expert is an individual network with expertise on a unique portion of the dataset, contributing to the collective capacity. Given a sample, delegator selects an expert and simultaneously outputs a rough prediction to trigger potential early termination. For each model in CoE, we propose a novel training algorithm with two major components: weight generation module (WGM) and label generation module (LGM). It fulfills the co-adaptation of experts and delegator. WGM partitions the training data into portions based on delegator via solving a balanced transportation problem, then impels each expert to focus on one portion by reweighting the losses. LGM generates the label to constitute the loss of delegator for expert selection. CoE achieves the state-of-the-art performance on ImageNet, 80.7% top-1 accuracy with 194M FLOPs. Combined with PWLU and CondConv, CoE further boosts the accuracy to 80.0% with only 100M FLOPs for the first time. Furthermore, experiment results on the translation task also demonstrate the strong generalizability of CoE. CoE is hardware-friendly, yielding a 3 6x acceleration compared with existing conditional computation approaches.} }
Endnote
%0 Conference Paper %T Collaboration of Experts: Achieving 80% Top-1 Accuracy on ImageNet with 100M FLOPs %A Yikang Zhang %A Zhuo Chen %A Zhao Zhong %B Proceedings of the 39th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2022 %E Kamalika Chaudhuri %E Stefanie Jegelka %E Le Song %E Csaba Szepesvari %E Gang Niu %E Sivan Sabato %F pmlr-v162-zhang22c %I PMLR %P 26068--26084 %U https://proceedings.mlr.press/v162/zhang22c.html %V 162 %X In this paper, we propose a Collaboration of Experts (CoE) framework to assemble the expertise of multiple networks towards a common goal. Each expert is an individual network with expertise on a unique portion of the dataset, contributing to the collective capacity. Given a sample, delegator selects an expert and simultaneously outputs a rough prediction to trigger potential early termination. For each model in CoE, we propose a novel training algorithm with two major components: weight generation module (WGM) and label generation module (LGM). It fulfills the co-adaptation of experts and delegator. WGM partitions the training data into portions based on delegator via solving a balanced transportation problem, then impels each expert to focus on one portion by reweighting the losses. LGM generates the label to constitute the loss of delegator for expert selection. CoE achieves the state-of-the-art performance on ImageNet, 80.7% top-1 accuracy with 194M FLOPs. Combined with PWLU and CondConv, CoE further boosts the accuracy to 80.0% with only 100M FLOPs for the first time. Furthermore, experiment results on the translation task also demonstrate the strong generalizability of CoE. CoE is hardware-friendly, yielding a 3 6x acceleration compared with existing conditional computation approaches.
APA
Zhang, Y., Chen, Z. & Zhong, Z.. (2022). Collaboration of Experts: Achieving 80% Top-1 Accuracy on ImageNet with 100M FLOPs. Proceedings of the 39th International Conference on Machine Learning, in Proceedings of Machine Learning Research 162:26068-26084 Available from https://proceedings.mlr.press/v162/zhang22c.html.

Related Material