ICML 2020poster17 citations

Orthogonalized SGD and Nested Architectures for Anytime Neural Networks

Chengcheng Wan, Henry Hoffmann, Shan Lu, Michael Maire

Abstract

We propose a novel variant of SGD customized for training network architectures that support anytime behavior: such networks produce a series of increasingly accurate outputs over time. Efficient architectural designs for these networks focus on re-using internal state; subnetworks must produce representations relevant for both imme- diate prediction as well as refinement by subse- quent network stages. We consider traditional branched networks as well as a new class of re- cursively nested networks. Our new optimizer, Orthogonalized SGD, dynamically re-balances task-specific gradients when training a multitask network. In the context of anytime architectures, this optimizer projects gradients from later out- puts onto a parameter subspace that does not in- terfere with those from earlier outputs. Experi- ments demonstrate that training with Orthogonal- ized SGD significantly improves generalization accuracy of anytime networks.

BibTeX
@InProceedings{pmlr-v119-wan20a,
  title = 	 {Orthogonalized {SGD} and Nested Architectures for Anytime Neural Networks},
  author =       {Wan, Chengcheng and Hoffmann, Henry and Lu, Shan and Maire, Michael},
  booktitle = 	 {Proceedings of the 37th International Conference on Machine Learning},
  pages = 	 {9807--9817},
  year = 	 {2020},
  editor = 	 {III, Hal Daumé and Singh, Aarti},
  volume = 	 {119},
  series = 	 {Proceedings of Machine Learning Research},
  month = 	 {13--18 Jul},
  publisher =    {PMLR},
  pdf = 	 {http://proceedings.mlr.press/v119/wan20a/wan20a.pdf},
  url = 	 {https://proceedings.mlr.press/v119/wan20a.html},
  abstract = 	 {We propose a novel variant of SGD customized for training network architectures that support anytime behavior: such networks produce a series of increasingly accurate outputs over time. Efficient architectural designs for these networks focus on re-using internal state; subnetworks must produce representations relevant for both imme- diate prediction as well as refinement by subse- quent network stages. We consider traditional branched networks as well as a new class of re- cursively nested networks. Our new optimizer, Orthogonalized SGD, dynamically re-balances task-specific gradients when training a multitask network. In the context of anytime architectures, this optimizer projects gradients from later out- puts onto a parameter subspace that does not in- terfere with those from earlier outputs. Experi- ments demonstrate that training with Orthogonal- ized SGD significantly improves generalization accuracy of anytime networks.}
}
Orthogonalized SGD and Nested Architectures for Anytime Neural Networks · ICML 2020