Timezone: »
Residual networks (ResNet) and weight normalization play an important role in various deep learning applications. However, parameter initialization strategies have not been studied previously for weight normalized networks and, in practice, initialization methods designed for un-normalized networks are used as a proxy. Similarly, initialization for ResNets have also been studied for un-normalized networks and often under simplified settings ignoring the shortcut connection. To address these issues, we propose a novel parameter initialization strategy that avoids explosion/vanishment of information across layers for weight normalized networks with and without residual connections. The proposed strategy is based on a theoretical analysis using mean field approximation. We run over 2,500 experiments and evaluate our proposal on image datasets showing that the proposed initialization outperforms existing initialization methods in terms of generalization performance, robustness to hyper-parameter values and variance between seeds, especially when networks get deeper in which case existing methods fail to even start training. Finally, we show that using our initialization in conjunction with learning rate warmup is able to reduce the gap between the performance of weight normalized and batch normalized networks.
Author Information
Devansh Arpit (Salesforce/MILA)
Víctor Campos (Barcelona Supercomputing Center)
Yoshua Bengio (Mila - University of Montreal)
More from the Same Authors
-
2022 Spotlight: Ensemble of Averages: Improving Model Selection and Boosting Performance in Domain Generalization »
Devansh Arpit · Huan Wang · Yingbo Zhou · Caiming Xiong -
2022 Spotlight: Lightning Talks 5B-1 »
Devansh Arpit · Xiaojun Xu · Zifan Shi · Ivan Skorokhodov · Shayan Shekarforoush · Zhan Tong · Yiqun Wang · Shichong Peng · Linyi Li · Ivan Skorokhodov · Huan Wang · Yibing Song · David Lindell · Yinghao Xu · Seyed Alireza Moazenipourasil · Sergey Tulyakov · Peter Wonka · Yiqun Wang · Ke Li · David Fleet · Yujun Shen · Yingbo Zhou · Bo Li · Jue Wang · Peter Wonka · Marcus Brubaker · Caiming Xiong · Limin Wang · Deli Zhao · Qifeng Chen · Dit-Yan Yeung -
2022 Poster: Ensemble of Averages: Improving Model Selection and Boosting Performance in Domain Generalization »
Devansh Arpit · Huan Wang · Yingbo Zhou · Caiming Xiong -
2019 : Climate Change: A Grand Challenge for ML »
Yoshua Bengio · Carla Gomes · Andrew Ng · Jeff Dean · Lester Mackey -
2019 Poster: Variational Temporal Abstraction »
Taesup Kim · Sungjin Ahn · Yoshua Bengio -
2018 : Poster Session »
Sujay Sanghavi · Vatsal Shah · Yanyao Shen · Tianchen Zhao · Yuandong Tian · Tomer Galanti · Mufan Li · Gilad Cohen · Daniel Rothchild · Aristide Baratin · Devansh Arpit · Vagelis Papalexakis · Michael Perlmutter · Ashok Vardhan Makkuva · Pim de Haan · Yingyan Lin · Wanmo Kang · Cheolhyoung Lee · Hao Shen · Sho Yaida · Dan Roberts · Nadav Cohen · Philippe Casgrain · Dejiao Zhang · Tengyu Ma · Avinash Ravichandran · Julian Emilio Salazar · Bo Li · Davis Liang · Christopher Wong · Glen Bigan Mbeng · Animesh Garg -
2018 : Opening remarks »
Yoshua Bengio -
2018 Poster: Image-to-image translation for cross-domain disentanglement »
Abel Gonzalez-Garcia · Joost van de Weijer · Yoshua Bengio -
2018 Poster: MetaGAN: An Adversarial Approach to Few-Shot Learning »
Ruixiang ZHANG · Tong Che · Zoubin Ghahramani · Yoshua Bengio · Yangqiu Song -
2018 Poster: Bayesian Model-Agnostic Meta-Learning »
Jaesik Yoon · Taesup Kim · Ousmane Dia · Sungwoong Kim · Yoshua Bengio · Sungjin Ahn -
2018 Poster: Sparse Attentive Backtracking: Temporal Credit Assignment Through Reminding »
Nan Rosemary Ke · Anirudh Goyal · Olexa Bilaniuk · Jonathan Binas · Michael Mozer · Chris Pal · Yoshua Bengio -
2018 Spotlight: Sparse Attentive Backtracking: Temporal Credit Assignment Through Reminding »
Nan Rosemary Ke · Anirudh Goyal · Olexa Bilaniuk · Jonathan Binas · Michael Mozer · Chris Pal · Yoshua Bengio -
2018 Spotlight: Bayesian Model-Agnostic Meta-Learning »
Jaesik Yoon · Taesup Kim · Ousmane Dia · Sungwoong Kim · Yoshua Bengio · Sungjin Ahn -
2018 Poster: Dendritic cortical microcircuits approximate the backpropagation algorithm »
João Sacramento · Rui Ponte Costa · Yoshua Bengio · Walter Senn -
2018 Oral: Dendritic cortical microcircuits approximate the backpropagation algorithm »
João Sacramento · Rui Ponte Costa · Yoshua Bengio · Walter Senn -
2017 : Yoshua Bengio »
Yoshua Bengio -
2017 : More Steps towards Biologically Plausible Backprop »
Yoshua Bengio -
2017 : A3T: Adversarially Augmented Adversarial Training »
Aristide Baratin · Simon Lacoste-Julien · Yoshua Bengio · Akram Erraqabi -
2017 : Competition III: The Conversational Intelligence Challenge »
Mikhail Burtsev · Ryan Lowe · Iulian Vlad Serban · Yoshua Bengio · Alexander Rudnicky · Alan W Black · Shrimai Prabhumoye · Artem Rodichev · Nikita Smetanin · Denis Fedorenko · CheongAn Lee · EUNMI HONG · Hwaran Lee · Geonmin Kim · Nicolas Gontier · Atsushi Saito · Andrey Gershfeld · Artem Burachenok -
2017 Poster: Variational Walkback: Learning a Transition Operator as a Stochastic Recurrent Net »
Anirudh Goyal · Nan Rosemary Ke · Surya Ganguli · Yoshua Bengio -
2017 Demonstration: A Deep Reinforcement Learning Chatbot »
Iulian Vlad Serban · Chinnadhurai Sankar · Mathieu Germain · Saizheng Zhang · Zhouhan Lin · Sandeep Subramanian · Taesup Kim · Michael Pieper · Sarath Chandar · Nan Rosemary Ke · Sai Rajeswar Mudumba · Alexandre de Brébisson · Jose Sotelo · Dendi A Suhubdy · Vincent Michalski · Joelle Pineau · Yoshua Bengio -
2017 Poster: GibbsNet: Iterative Adversarial Inference for Deep Graphical Models »
Alex Lamb · R Devon Hjelm · Yaroslav Ganin · Joseph Paul Cohen · Aaron Courville · Yoshua Bengio -
2017 Poster: Plan, Attend, Generate: Planning for Sequence-to-Sequence Models »
Caglar Gulcehre · Francis Dutil · Adam Trischler · Yoshua Bengio -
2017 Poster: Z-Forcing: Training Stochastic Recurrent Networks »
Anirudh Goyal · Alessandro Sordoni · Marc-Alexandre Côté · Nan Rosemary Ke · Yoshua Bengio