Timezone: »
Poster
Hidden Progress in Deep Learning: SGD Learns Parities Near the Computational Limit
Boaz Barak · Benjamin Edelman · Surbhi Goel · Sham Kakade · Eran Malach · Cyril Zhang
There is mounting evidence of emergent phenomena in the capabilities of deep learning methods as we scale up datasets, model sizes, and training times. While there are some accounts of how these resources modulate statistical capacity, far less is known about their effect on the computational problem of model training. This work conducts such an exploration through the lens of learning a $k$-sparse parity of $n$ bits, a canonical discrete search problem which is statistically easy but computationally hard. Empirically, we find that a variety of neural networks successfully learn sparse parities, with discontinuous phase transitions in the training curves. On small instances, learning abruptly occurs at approximately $n^{O(k)}$ iterations; this nearly matches SQ lower bounds, despite the apparent lack of a sparse prior. Our theoretical analysis shows that these observations are not explained by a Langevin-like mechanism, whereby SGD "stumbles in the dark" until it finds the hidden set of features (a natural algorithm which also runs in $n^{O(k)}$ time). Instead, we show that SGD gradually amplifies the sparse solution via a Fourier gap in the population gradient, making continual progress that is invisible to loss and error metrics.
Author Information
Boaz Barak (Harvard University)
Benjamin Edelman (Harvard University)
Surbhi Goel (Microsoft Research NYC)
Sham Kakade (Harvard University & Amazon)
Eran Malach (Hebrew University Jerusalem Israel)
Cyril Zhang (Microsoft Research NYC)
More from the Same Authors
-
2021 Spotlight: On the Power of Differentiable Learning versus PAC and SQ Learning »
Emmanuel Abbe · Pritish Kamath · Eran Malach · Colin Sandon · Nathan Srebro -
2021 : Catastrophic Failures of Neural Active Learning on Heteroskedastic Distributions »
Savya Khosla · Alex Lamb · Jordan Ash · Cyril Zhang · Kenji Kawaguchi -
2022 : The Trade-offs of Incremental Linearization Algorithms for Nonsmooth Composite Problems »
Krishna Pillutla · Vincent Roulet · Sham Kakade · Zaid Harchaoui -
2022 : Deconstructing Distributions: A Pointwise Framework of Learning »
Gal Kaplun · Nikhil Ghosh · Saurabh Garg · Boaz Barak · Preetum Nakkiran -
2023 Workshop: Mathematics of Modern Machine Learning (M3L) »
Aditi Raghunathan · Alex Damian · Bingbin Liu · Christina Baek · Kaifeng Lyu · Surbhi Goel · Tengyu Ma · Zhiyuan Li -
2022 : Matryoshka Representations for Adaptive Deployment »
Aniket Rege · Aditya Kusupati · Gantavya Bhatt · Matthew Wallingford · Aditya Sinha · Vivek Ramanujan · William Howard-Snyder · Kaifeng Chen · Sham Kakade · Prateek Jain · Ali Farhadi -
2022 Spotlight: Risk Bounds of Multi-Pass SGD for Least Squares in the Interpolation Regime »
Difan Zou · Jingfeng Wu · Vladimir Braverman · Quanquan Gu · Sham Kakade -
2022 Poster: Recurrent Convolutional Neural Networks Learn Succinct Learning Algorithms »
Surbhi Goel · Sham Kakade · Adam Kalai · Cyril Zhang -
2022 Poster: The Power and Limitation of Pretraining-Finetuning for Linear Regression under Covariate Shift »
Jingfeng Wu · Difan Zou · Vladimir Braverman · Quanquan Gu · Sham Kakade -
2022 Poster: Unpacking Reward Shaping: Understanding the Benefits of Reward Engineering on Sample Complexity »
Abhishek Gupta · Aldo Pacchiano · Yuexiang Zhai · Sham Kakade · Sergey Levine -
2022 Poster: Risk Bounds of Multi-Pass SGD for Least Squares in the Interpolation Regime »
Difan Zou · Jingfeng Wu · Vladimir Braverman · Quanquan Gu · Sham Kakade -
2022 Poster: Knowledge Distillation: Bad Models Can Be Good Role Models »
Gal Kaplun · Eran Malach · Preetum Nakkiran · Shai Shalev-Shwartz -
2022 Poster: Matryoshka Representation Learning »
Aditya Kusupati · Gantavya Bhatt · Aniket Rege · Matthew Wallingford · Aditya Sinha · Vivek Ramanujan · William Howard-Snyder · Kaifeng Chen · Sham Kakade · Prateek Jain · Ali Farhadi -
2021 : Invited Speaker Panel »
Sham Kakade · Minmin Chen · Philip Thomas · Angela Schoellig · Barbara Engelhardt · Doina Precup · George Tucker -
2021 : Q&A for Sham Kakade »
Sham Kakade -
2021 : Generalization theory in Offline RL »
Sham Kakade -
2021 Poster: On the Power of Differentiable Learning versus PAC and SQ Learning »
Emmanuel Abbe · Pritish Kamath · Eran Malach · Colin Sandon · Nathan Srebro -
2021 Poster: The Benefits of Implicit Regularization from SGD in Least Squares Problems »
Difan Zou · Jingfeng Wu · Vladimir Braverman · Quanquan Gu · Dean Foster · Sham Kakade -
2021 Poster: Robust and differentially private mean estimation »
Xiyang Liu · Weihao Kong · Sham Kakade · Sewoong Oh -
2021 Poster: An Exponential Lower Bound for Linearly Realizable MDP with Constant Suboptimality Gap »
Yuanhao Wang · Ruosong Wang · Sham Kakade -
2021 Poster: Going Beyond Linear RL: Sample Efficient Neural Function Approximation »
Baihe Huang · Kaixuan Huang · Sham Kakade · Jason Lee · Qi Lei · Runzhe Wang · Jiaqi Yang -
2021 Poster: LLC: Accurate, Multi-purpose Learnt Low-dimensional Binary Codes »
Aditya Kusupati · Matthew Wallingford · Vivek Ramanujan · Raghav Somani · Jae Sung Park · Krishna Pillutla · Prateek Jain · Sham Kakade · Ali Farhadi -
2021 Poster: Gone Fishing: Neural Active Learning with Fisher Embeddings »
Jordan Ash · Surbhi Goel · Akshay Krishnamurthy · Sham Kakade -
2021 Poster: Revisiting Model Stitching to Compare Neural Representations »
Yamini Bansal · Preetum Nakkiran · Boaz Barak -
2021 Poster: Optimal Gradient-based Algorithms for Non-concave Bandit Optimization »
Baihe Huang · Kaixuan Huang · Sham Kakade · Jason Lee · Qi Lei · Runzhe Wang · Jiaqi Yang -
2021 Oral: An Exponential Lower Bound for Linearly Realizable MDP with Constant Suboptimality Gap »
Yuanhao Wang · Ruosong Wang · Sham Kakade -
2020 Tutorial: (Track3) Policy Optimization in Reinforcement Learning Q&A »
Sham M Kakade · Martha White · Nicolas Le Roux -
2020 Poster: The Implications of Local Correlation on Learning Some Deep Functions »
Eran Malach · Shai Shalev-Shwartz -
2020 Poster: Robust Meta-learning for Mixed Linear Regression with Small Batches »
Weihao Kong · Raghav Somani · Sham Kakade · Sewoong Oh -
2020 Poster: Is Long Horizon RL More Difficult Than Short Horizon RL? »
Ruosong Wang · Simon Du · Lin Yang · Sham Kakade -
2020 Poster: Learning Parities with Neural Networks »
Amit Daniely · Eran Malach -
2020 Poster: FLAMBE: Structural Complexity and Representation Learning of Low Rank MDPs »
Alekh Agarwal · Sham Kakade · Akshay Krishnamurthy · Wen Sun -
2020 Poster: PC-PG: Policy Cover Directed Exploration for Provable Policy Gradient Learning »
Alekh Agarwal · Mikael Henaff · Sham Kakade · Wen Sun -
2020 Poster: Sample-Efficient Reinforcement Learning of Undercomplete POMDPs »
Chi Jin · Sham Kakade · Akshay Krishnamurthy · Qinghua Liu -
2020 Spotlight: Sample-Efficient Reinforcement Learning of Undercomplete POMDPs »
Chi Jin · Sham Kakade · Akshay Krishnamurthy · Qinghua Liu -
2020 Oral: Learning Parities with Neural Networks »
Amit Daniely · Eran Malach -
2020 Oral: FLAMBE: Structural Complexity and Representation Learning of Low Rank MDPs »
Alekh Agarwal · Sham Kakade · Akshay Krishnamurthy · Wen Sun -
2020 Poster: Model-Based Multi-Agent RL in Zero-Sum Markov Games with Near-Optimal Sample Complexity »
Kaiqing Zhang · Sham Kakade · Tamer Basar · Lin Yang -
2020 Poster: Information Theoretic Regret Bounds for Online Nonlinear Control »
Sham Kakade · Akshay Krishnamurthy · Kendall Lowrey · Motoya Ohnishi · Wen Sun -
2020 Spotlight: Model-Based Multi-Agent RL in Zero-Sum Markov Games with Near-Optimal Sample Complexity »
Kaiqing Zhang · Sham Kakade · Tamer Basar · Lin Yang -
2020 Tutorial: (Track3) Policy Optimization in Reinforcement Learning »
Sham M Kakade · Martha White · Nicolas Le Roux -
2019 : Poster and Coffee Break 2 »
Karol Hausman · Kefan Dong · Ken Goldberg · Lihong Li · Lin Yang · Lingxiao Wang · Lior Shani · Liwei Wang · Loren Amdahl-Culleton · Lucas Cassano · Marc Dymetman · Marc Bellemare · Marcin Tomczak · Margarita Castro · Marius Kloft · Marius-Constantin Dinu · Markus Holzleitner · Martha White · Mengdi Wang · Michael Jordan · Mihailo Jovanovic · Ming Yu · Minshuo Chen · Moonkyung Ryu · Muhammad Zaheer · Naman Agarwal · Nan Jiang · Niao He · Nikolaus Yasui · Nikos Karampatziakis · Nino Vieillard · Ofir Nachum · Olivier Pietquin · Ozan Sener · Pan Xu · Parameswaran Kamalaruban · Paul Mineiro · Paul Rolland · Philip Amortila · Pierre-Luc Bacon · Prakash Panangaden · Qi Cai · Qiang Liu · Quanquan Gu · Raihan Seraj · Richard Sutton · Rick Valenzano · Robert Dadashi · Rodrigo Toro Icarte · Roshan Shariff · Roy Fox · Ruosong Wang · Saeed Ghadimi · Samuel Sokota · Sean Sinclair · Sepp Hochreiter · Sergey Levine · Sergio Valcarcel Macua · Sham Kakade · Shangtong Zhang · Sheila McIlraith · Shie Mannor · Shimon Whiteson · Shuai Li · Shuang Qiu · Wai Lok Li · Siddhartha Banerjee · Sitao Luan · Tamer Basar · Thinh Doan · Tianhe Yu · Tianyi Liu · Tom Zahavy · Toryn Klassen · Tuo Zhao · Vicenç Gómez · Vincent Liu · Volkan Cevher · Wesley Suttle · Xiao-Wen Chang · Xiaohan Wei · Xiaotong Liu · Xingguo Li · Xinyi Chen · Xingyou Song · Yao Liu · YiDing Jiang · Yihao Feng · Yilun Du · Yinlam Chow · Yinyu Ye · Yishay Mansour · · Yonathan Efroni · Yongxin Chen · Yuanhao Wang · Bo Dai · Chen-Yu Wei · Harsh Shrivastava · Hongyang Zhang · Qinqing Zheng · SIDDHARTHA SATPATHI · Xueqing Liu · Andreu Vall -
2019 : Poster Spotlight 2 »
Aaron Sidford · Mengdi Wang · Lin Yang · Yinyu Ye · Zuyue Fu · Zhuoran Yang · Yongxin Chen · Zhaoran Wang · Ofir Nachum · Bo Dai · Ilya Kostrikov · Dale Schuurmans · Ziyang Tang · Yihao Feng · Lihong Li · Denny Zhou · Qiang Liu · Rodrigo Toro Icarte · Ethan Waldie · Toryn Klassen · Rick Valenzano · Margarita Castro · Simon Du · Sham Kakade · Ruosong Wang · Minshuo Chen · Tianyi Liu · Xingguo Li · Zhaoran Wang · Tuo Zhao · Philip Amortila · Doina Precup · Prakash Panangaden · Marc Bellemare -
2019 : The Provable Effectiveness of Policy Gradient Methods in Reinforcement Learning »
Sham Kakade -
2019 Poster: The Step Decay Schedule: A Near Optimal, Geometrically Decaying Learning Rate Procedure For Least Squares »
Rong Ge · Sham Kakade · Rahul Kidambi · Praneeth Netrapalli -
2019 Poster: SGD on Neural Networks Learns Functions of Increasing Complexity »
Dimitris Kalimeris · Gal Kaplun · Preetum Nakkiran · Benjamin Edelman · Tristan Yang · Boaz Barak · Haofeng Zhang -
2019 Spotlight: SGD on Neural Networks Learns Functions of Increasing Complexity »
Dimitris Kalimeris · Gal Kaplun · Preetum Nakkiran · Benjamin Edelman · Tristan Yang · Boaz Barak · Haofeng Zhang -
2019 Poster: Is Deeper Better only when Shallow is Good? »
Eran Malach · Shai Shalev-Shwartz -
2019 Poster: Meta-Learning with Implicit Gradients »
Aravind Rajeswaran · Chelsea Finn · Sham Kakade · Sergey Levine -
2019 Poster: (Nearly) Efficient Algorithms for the Graph Matching Problem on Correlated Random Graphs »
Boaz Barak · Chi-Ning Chou · Zhixian Lei · Tselil Schramm · Yueqi Sheng -
2018 Poster: A Smoother Way to Train Structured Prediction Models »
Krishna Pillutla · Vincent Roulet · Sham Kakade · Zaid Harchaoui -
2018 Poster: Provably Correct Automatic Sub-Differentiation for Qualified Programs »
Sham Kakade · Jason Lee -
2017 Poster: Learning Overcomplete HMMs »
Vatsal Sharan · Sham Kakade · Percy Liang · Gregory Valiant -
2017 Poster: Eigenvalue Decay Implies Polynomial-Time Learnability for Neural Networks »
Surbhi Goel · Adam Klivans -
2017 Poster: Decoupling "when to update" from "how to update" »
Eran Malach · Shai Shalev-Shwartz -
2017 Poster: Towards Generalization and Simplicity in Continuous Control »
Aravind Rajeswaran · Kendall Lowrey · Emanuel Todorov · Sham Kakade -
2016 Poster: Provable Efficient Online Matrix Completion via Non-convex Stochastic Gradient Descent »
Chi Jin · Sham Kakade · Praneeth Netrapalli -
2015 Poster: Convergence Rates of Active Learning for Maximum Likelihood Estimation »
Kamalika Chaudhuri · Sham Kakade · Praneeth Netrapalli · Sujay Sanghavi -
2015 Poster: Super-Resolution Off the Grid »
Qingqing Huang · Sham Kakade -
2015 Spotlight: Super-Resolution Off the Grid »
Qingqing Huang · Sham Kakade -
2013 Poster: When are Overcomplete Topic Models Identifiable? Uniqueness of Tensor Tucker Decompositions with Structured Sparsity »
Anima Anandkumar · Daniel Hsu · Majid Janzamin · Sham M Kakade -
2012 Poster: Learning Mixtures of Tree Graphical Models »
Anima Anandkumar · Daniel Hsu · Furong Huang · Sham M Kakade -
2012 Poster: A Spectral Algorithm for Latent Dirichlet Allocation »
Anima Anandkumar · Dean P Foster · Daniel Hsu · Sham M Kakade · Yi-Kai Liu -
2012 Poster: Identifiability and Unmixing of Latent Parse Trees »
Percy Liang · Sham M Kakade · Daniel Hsu -
2012 Spotlight: A Spectral Algorithm for Latent Dirichlet Allocation »
Anima Anandkumar · Dean P Foster · Daniel Hsu · Sham M Kakade · Yi-Kai Liu -
2011 Poster: Stochastic convex optimization with bandit feedback »
Alekh Agarwal · Dean P Foster · Daniel Hsu · Sham M Kakade · Sasha Rakhlin -
2011 Poster: Spectral Methods for Learning Multivariate Latent Tree Structure »
Anima Anandkumar · Kamalika Chaudhuri · Daniel Hsu · Sham M Kakade · Le Song · Tong Zhang -
2011 Poster: Efficient Learning of Generalized Linear and Single Index Models with Isotonic Regression »
Sham M Kakade · Adam Kalai · Varun Kanade · Ohad Shamir -
2010 Spotlight: Learning from Logged Implicit Exploration Data »
Alex Strehl · Lihong Li · John Langford · Sham M Kakade -
2010 Poster: Learning from Logged Implicit Exploration Data »
Alexander L Strehl · John Langford · Lihong Li · Sham M Kakade -
2009 Poster: Multi-Label Prediction via Compressed Sensing »
Daniel Hsu · Sham M Kakade · John Langford · Tong Zhang -
2009 Oral: Multi-Label Prediction via Compressed Sensing »
Daniel Hsu · Sham M Kakade · John Langford · Tong Zhang -
2008 Poster: Mind the Duality Gap: Logarithmic regret algorithms for online optimization »
Shai Shalev-Shwartz · Sham M Kakade -
2008 Poster: On the Generalization Ability of Online Strongly Convex Programming Algorithms »
Sham M Kakade · Ambuj Tewari -
2008 Spotlight: On the Generalization Ability of Online Strongly Convex Programming Algorithms »
Sham M Kakade · Ambuj Tewari -
2008 Spotlight: Mind the Duality Gap: Logarithmic regret algorithms for online optimization »
Shai Shalev-Shwartz · Sham M Kakade -
2008 Poster: On the Complexity of Linear Prediction: Risk Bounds, Margin Bounds, and Regularization »
Sham M Kakade · Karthik Sridharan · Ambuj Tewari -
2007 Oral: The Price of Bandit Information for Online Optimization »
Varsha Dani · Thomas P Hayes · Sham M Kakade -
2007 Poster: The Price of Bandit Information for Online Optimization »
Varsha Dani · Thomas P Hayes · Sham M Kakade