Timezone: »
Inspired by BatchNorm, there has been an explosion of normalization layers in deep learning. Recent works have identified a multitude of beneficial properties in BatchNorm to explain its success. However, given the pursuit of alternative normalization layers, these properties need to be generalized so that any given layer's success/failure can be accurately predicted. In this work, we take a first step towards this goal by extending known properties of BatchNorm in randomly initialized deep neural networks (DNNs) to several recently proposed normalization layers. Our primary findings follow: (i) similar to BatchNorm, activations-based normalization layers can prevent exponential growth of activations in ResNets, but parametric techniques require explicit remedies; (ii) use of GroupNorm can ensure an informative forward propagation, with different samples being assigned dissimilar activations, but increasing group size results in increasingly indistinguishable activations for different samples, explaining slow convergence speed in models with LayerNorm; and (iii) small group sizes result in large gradient norm in earlier layers, hence explaining training instability issues in Instance Normalization and illustrating a speed-stability tradeoff in GroupNorm. Overall, our analysis reveals a unified set of mechanisms that underpin the success of normalization methods in deep learning, providing us with a compass to systematically explore the vast design space of DNN normalization layers.
Author Information
Ekdeep S Lubana (University of Michigan)
Robert Dick (University of Michigan)
Hidenori Tanaka (NTT Research, PHI Lab / Stanford University)
More from the Same Authors
-
2022 : Mechanistic Lens on Mode Connectivity »
Ekdeep S Lubana · Eric Bigelow · Robert Dick · David Krueger · Hidenori Tanaka -
2022 : What shapes the loss landscape of self-supervised learning? »
Liu Ziyin · Ekdeep S Lubana · Masahito Ueda · Hidenori Tanaka -
2022 : Geometric Considerations for Normalization Layers in Equivariant Neural Networks »
Max Aalto · Ekdeep S Lubana · Hidenori Tanaka -
2022 : A Mechanistic Lens on Mode Connectivity »
Ekdeep S Lubana · Eric Bigelow · Robert Dick · David Krueger · Hidenori Tanaka -
2022 Poster: Analyzing Data-Centric Properties for Graph Contrastive Learning »
Puja Trivedi · Ekdeep S Lubana · Mark Heimann · Danai Koutra · Jayaraman Thiagarajan -
2021 Poster: Noether’s Learning Dynamics: Role of Symmetry Breaking in Neural Networks »
Hidenori Tanaka · Daniel Kunin -
2020 Poster: Pruning neural networks without any data by iteratively conserving synaptic flow »
Hidenori Tanaka · Daniel Kunin · Daniel Yamins · Surya Ganguli -
2019 : Poster Session »
Pravish Sainath · Mohamed Akrout · Charles Delahunt · Nathan Kutz · Guangyu Robert Yang · Joseph Marino · L F Abbott · Nicolas Vecoven · Damien Ernst · andrew warrington · Michael Kagan · Kyunghyun Cho · Kameron Harris · Leopold Grinberg · John J. Hopfield · Dmitry Krotov · Taliah Muhammad · Erick Cobos · Edgar Walker · Jacob Reimer · Andreas Tolias · Alexander Ecker · Janaki Sheth · Yu Zhang · Maciej Wołczyk · Jacek Tabor · Szymon Maszke · Roman Pogodin · Dane Corneil · Wulfram Gerstner · Baihan Lin · Guillermo Cecchi · Jenna M Reinen · Irina Rish · Guillaume Bellec · Darjan Salaj · Anand Subramoney · Wolfgang Maass · Yueqi Wang · Ari Pakman · Jin Hyung Lee · Liam Paninski · Bryan Tripp · Colin Graber · Alex Schwing · Luke Prince · Gabriel Ocker · Michael Buice · Benjamin Lansdell · Konrad Kording · Jack Lindsey · Terrence Sejnowski · Matthew Farrell · Eric Shea-Brown · Nicolas Farrugia · Victor Nepveu · Jiwoong Im · Kristin Branson · Brian Hu · Ramakrishnan Iyer · Stefan Mihalas · Sneha Aenugu · Hananel Hazan · Sihui Dai · Tan Nguyen · Doris Tsao · Richard Baraniuk · Anima Anandkumar · Hidenori Tanaka · Aran Nayebi · Stephen Baccus · Surya Ganguli · Dean Pospisil · Eilif Muller · Jeffrey S Cheng · Gaël Varoquaux · Kamalaker Dadi · Dimitrios C Gklezakos · Rajesh PN Rao · Anand Louis · Christos Papadimitriou · Santosh Vempala · Naganand Yadati · Daniel Zdeblick · Daniela M Witten · Nicholas Roberts · Vinay Prabhu · Pierre Bellec · Poornima Ramesh · Jakob H Macke · Santiago Cadena · Guillaume Bellec · Franz Scherr · Owen Marschall · Robert Kim · Hannes Rapp · Marcio Fonseca · Oliver Armitage · Jiwoong Im · Thomas Hardcastle · Abhishek Sharma · Wyeth Bair · Adrian Valente · Shane Shang · Merav Stern · Rutuja Patil · Peter Wang · Sruthi Gorantla · Peter Stratton · Tristan Edwards · Jialin Lu · Martin Ester · Yurii Vlasov · Siavash Golkar -
2019 Poster: From deep learning to mechanistic understanding in neuroscience: the structure of retinal prediction »
Hidenori Tanaka · Aran Nayebi · Niru Maheswaranathan · Lane McIntosh · Stephen Baccus · Surya Ganguli