Timezone: »
Knowledge distillation introduced in the deep learning context is a method to transfer knowledge from one architecture to another. In particular, when the architectures are identical, this is called self-distillation. The idea is to feed in predictions of the trained model as new target values for retraining (and iterate this loop possibly a few times). It has been empirically observed that the self-distilled model often achieves higher accuracy on held out data. Why this happens, however, has been a mystery: the self-distillation dynamics does not receive any new information about the task and solely evolves by looping over training. To the best of our knowledge, there is no rigorous understanding of why this happens. This work provides the first theoretical analysis of self-distillation. We focus on fitting a nonlinear function to training data, where the model space is Hilbert space and fitting is subject to L2 regularization in this function space. We show that self-distillation iterations modify regularization by progressively limiting the number of basis functions that can be used to represent the solution. This implies (as we also verify empirically) that while a few rounds of self-distillation may reduce over-fitting, further rounds may lead to under-fitting and thus worse performance.
Author Information
Hossein Mobahi (Google Research)
Mehrdad Farajtabar (DeepMind)
Peter Bartlett (UC Berkeley)
More from the Same Authors
-
2022 : Trajectory ensembling for fine tuning - performance gains without modifying training »
Louise Anderson-Conway · Vighnesh Birodkar · Saurabh Singh · Hossein Mobahi · Alexander Alemi -
2020 Poster: Learning to Incentivize Other Learning Agents »
Jiachen Yang · Ang Li · Mehrdad Farajtabar · Peter Sunehag · Edward Hughes · Hongyuan Zha -
2020 Poster: Understanding the Role of Training Regimes in Continual Learning »
Seyed Iman Mirzadeh · Mehrdad Farajtabar · Razvan Pascanu · Hassan Ghasemzadeh -
2020 Session: Orals & Spotlights Track 17: Kernel Methods/Optimization »
Chiranjib Bhattacharyya · Hossein Mobahi -
2020 Poster: A Maximum-Entropy Approach to Off-Policy Evaluation in Average-Reward MDPs »
Nevena Lazic · Dong Yin · Mehrdad Farajtabar · Nir Levine · Dilan Gorur · Chris Harris · Dale Schuurmans -
2019 : Contributed Session - Spotlight Talks »
Jonathan Frankle · David Schwab · Ari Morcos · Qianli Ma · Yao-Hung Hubert Tsai · Ruslan Salakhutdinov · YiDing Jiang · Dilip Krishnan · Hossein Mobahi · Samy Bengio · Sho Yaida · Muqiao Yang -
2019 : Lunch Break and Posters »
Xingyou Song · Elad Hoffer · Wei-Cheng Chang · Jeremy Cohen · Jyoti Islam · Yaniv Blumenfeld · Andreas Madsen · Jonathan Frankle · Sebastian Goldt · Satrajit Chatterjee · Abhishek Panigrahi · Alex Renda · Brian Bartoldson · Israel Birhane · Aristide Baratin · Niladri Chatterji · Roman Novak · Jessica Forde · YiDing Jiang · Yilun Du · Linara Adilova · Michael Kamp · Berry Weinstein · Itay Hubara · Tal Ben-Nun · Torsten Hoefler · Daniel Soudry · Hsiang-Fu Yu · Kai Zhong · Yiming Yang · Inderjit Dhillon · Jaime Carbonell · Yanqing Zhang · Dar Gilboa · Johannes Brandstetter · Alexander R Johansen · Gintare Karolina Dziugaite · Raghav Somani · Ari Morcos · Freddie Kalaitzis · Hanie Sedghi · Lechao Xiao · John Zech · Muqiao Yang · Simran Kaur · Qianli Ma · Yao-Hung Hubert Tsai · Ruslan Salakhutdinov · Sho Yaida · Zachary Lipton · Daniel Roy · Michael Carbin · Florent Krzakala · Lenka Zdeborová · Guy Gur-Ari · Ethan Dyer · Dilip Krishnan · Hossein Mobahi · Samy Bengio · Behnam Neyshabur · Praneeth Netrapalli · Kris Sankaran · Julien Cornebise · Yoshua Bengio · Vincent Michalski · Samira Ebrahimi Kahou · Md Rifat Arefin · Jiri Hron · Jaehoon Lee · Jascha Sohl-Dickstein · Samuel Schoenholz · David Schwab · Dongyu Li · Sang Keun Choe · Henning Petzka · Ashish Verma · Zhichao Lin · Cristian Sminchisescu -
2018 Poster: Large Margin Deep Networks for Classification »
Gamaleldin Elsayed · Dilip Krishnan · Hossein Mobahi · Kevin Regan · Samy Bengio