I work on theoretical aspects of machine learning. Recent topics include adaptive data analysis, complexity of learning with constrained access to data, and privacy-preserving learning. I also worked on understanding of natural learning systems: learning by the brain and evolution as learning.
I serve as a director and treasurer on the steering committee of the Association for Computational Learning
Recent program committees:
- Co-organizer of Privacy and the Science of Data Analysis workshop at the Simons Institute for the Theory of Computing, UC Berkeley. Apr 8-12, 2019 (with Kamalika Chaudhuri, Sesa Slavkovic, Anand Sarwate and Adam Smith).
- Co-organizer of Google-Simons Institute workshop on privacy. Apr. 5, 2019 (with Ravi Kumar, Ilya Mironov and Adam Smith).
- Co-organizer of Data Privacy: Foundations and Applications program at the Simons Institute for the Theory of Computing, UC Berkeley. Jan-May 2019 (with Katrina Ligett, Kobbi Nissim, Sesa Slavkovic and Adam Smith).
- July 2017: visiting Centrum Wiskunde & Informatica (CWI), Amsterdam (hosted by Peter Grunwald).
- May 2017: Co-organizer of Data Privacy: Planning Workshop at the Simons Institute.
- March-May 2017: visiting scientist at the Simons Institute. Foundations of Machine Learning program.
- Dec 2016: Co-organizer of workshop on Adaptive Data Analysis (WADAPT) at NIPS 2016 (with Aaditya Ramdas, Aaron Roth and Adam Smith).
- Interaction is necessary for distributed learning with privacy or communication constraints.
With Yuval Dagan. STOC 2020 .
- Private Stochastic Convex Optimization: Optimal Rates in Linear Time.
With Tomer Koren and Kunal Talwar. STOC 2020 .
- PAC learning with stable and private predictions .
With Yuval Dagan. Nov 2019 .
- Does Learning Require Memorization? A Short Tale about a Long Tail.
STOC 2020 .
- Private Stochastic Convex Optimization with Optimal Rates.
With Raef Bassily, Kunal Talwar and Abhradeep Thakurta. NeurIPS 2019 (spotlight). .
- High probability generalization bounds for uniformly stable algorithms with nearly optimal rate.
With Jan Vondrak. COLT 2019 .
- The advantages of multiple classes for reducing overfitting from test set reuse.
With Roy Frostig and Moritz Hardt. ICML 2019 .
- Locally Private Learning without Interaction Requires Separation.
With Amit Daniely. NeurIPS 2019.
- Amplification by Shuffling: From Local to Central Differential Privacy.
With Ulfar Erlingsson, Ilya Mironov, Ananth Raghunathan, Kunal Talwar, Abhradeep Thakurta. SODA 2019.
- Generalization Bounds for Uniformly Stable Algorithms.
With Jan Vondrak. NIPS 2018 (spotlight).
- Privacy Amplification by Iteration.
With Ilya Mironov, Kunal Talwar and Abhradeep Thakurta. FOCS 2018 .
- Privacy-preserving Prediction.
With Cynthia Dwork. COLT 2018 .
- Calibrating Noise to Variance in Adaptive Data Analysis.
With Thomas Steinke. COLT 2018 .
- A General Characterization of the Statistical Query Complexity.
COLT 2017 .
- Generalization of ERM in Stochastic Convex Optimization: The Dimension Strikes Back.
NIPS 2016 (oral presentation).
- The reusable holdout: Preserving validity in adaptive data analysis.
With Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold and Aaron Roth. Science, 2015.
IBM Research 2015 Best Paper Award.
Based on STOC and NIPS papers below. See also my post on this work at IBM Research blog (republished by KDnuggets).
- Generalization in Adaptive Data Analysis and Holdout Reuse.
With Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold and Aaron Roth. NIPS, 2015.
- Preserving Statistical Validity in Adaptive Data Analysis.
With Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold and Aaron Roth. STOC 2015.
Invited to SICOMP special issue on STOC
- On the Complexity of Random Satisfiability Problems with Planted Solutions .
With Will Perkins and Santosh Vempala. STOC 2015 .
- Sample Complexity Bounds on Differentially Private Learning via Communication Complexity
With David Xiao. COLT 2014, SICOMP 2015.
- Optimal Bounds on Approximation of Submodular and XOS Functions by Juntas.
With Jan Vondrak. FOCS 2013 . SICOMP 2016, Special issue on FOCS
IBM Research 2016 Best Paper Award.
- Learning using Local Membership Queries.
With Pranjal Awasthi and Varun Kanade. COLT 2013, Best Student (co-authored) Paper Award
- Statistical Algorithms and a Lower Bound for Detecting Planted Cliques.
With Elena Grigorescu, Lev Reyzin, Santosh Vempala and Ying Xiao. STOC 2013 . JACM 2017 .
- Nearly Optimal Solutions for the Chow Parameters Problem and Low-weight Approximation of Halfspaces.
With Anindya De, Ilias Diakonikolas and Rocco Servedio. STOC 2012; JACM 2014 .
IBM Research 2014 Best Paper Award.
- Distribution-Specific Agnostic Boosting.
ITCS (formerly ICS) 2010.
- A Complete Characterization of Statistical Query Learning with Applications to Evolvability.
FOCS 2009; JCSS 2012 (Special issue on Learning Theory).
- Experience-Induced Neural Circuits That Achieve High Capacity..
With Leslie Valiant. Neural Computation 21:10, 2009.
- New Results for Learning Noisy Parities and Halfspaces.
With Parikshit Gopalan, Subhash Khot, and Ashok Ponnuswami. FOCS 2006; SICOMP 2009, Special issue on FOCS
- Hardness of Approximate Two-level Logic Minimization and PAC Learning with Membership Queries.
STOC 2006; JCSS 75(1), 2009 (Special issue on Learning Theory)
- Attribute Efficient and Non-adaptive Learning of Parities and DNF Expressions.
COLT 2005, Best Student Paper Award; JMLR 2007, Special issue on COLT
- The Complexity of Properly Learning Simple Concept Classes.
With Misha Alekhnovich, Mark Braverman, Adam Klivans, and Toni Pitassi.
FOCS 2004; JCSS 74(1), 2008 (Special issue on Learning Theory)
- Does Learning Require Memorization? A Short Tale about a Long Tail. Princeton/Harvard/MIT/TAU/ATI/DeepMind/Google (Sep 2019+):slides, video courtesy of Alan Turing Institute.
- High probability generalization bounds for uniformly stable algorithms with nearly optimal rate. COLT 2019:slides, video
- Locally Private Learning without Interaction Requires Separation. Privacy and the Science of Data Analysis workshop (Apr 2019):slides.
- Amplification by Shuffling: From Local to Central Differential Privacy. Privacy-preserving Machine Learning workshop (Dec 2018); ITA 2019, Simons Institute seminar (Feb 2019):slides.
- Privacy-preserving prediction. COLT 2018; Privacy in Graphs workshop (Nov 2018); JSM 2019. slides.
- Generalization bounds for uniformly stable algorithms. Robust and High-Dimensional Statistics workshop (Oct 2018):slides, video. NIPS spotlight video.
- Stability, Information and Generalization in Adaptive Data Analysis. Google NYC/Princeton/Penn (Apr. 2018): slides.
- Dealing with Range Anxiety in Mean Estimation. ALT 2017 (Nov 2017): slides.
- A General Characterization of the Statistical Query Complexity. COLT 2017 (July 2017); NYU (Feb. 2018): slides.
- Understanding Generalization in Adaptive Data Analysis Computational Challenges in Machine Learning workshop, EPFL, and Bertinoro (2017):slides, video.
- On the power of learning from k-wise queries. ITCS 2017: slides, video.
- Lower bounds against convex relaxations via the statistical query complexity. Caltech/UCLA/Stanford/Harvard/MIT, 2017: slides (with some comments in the notes).
- Generalization of ERM in stochastic convex optimization. NIPS 2016: slides and video
- Generalization and adaptivity in stochastic convex optimization. TOCA-SV 2016: slides (with some comments in the notes).
- Generalization in Adaptive Data Analysis via Max-Information. Simons Institute workshop on Information Theory, 2016: slides.
- Preserving Validity in Adaptive Data Analysis. National Academy of Engineering, 2016: slides.
- Adaptive Data Analysis without Overfitting. Workshop on Learning. NUS, 2015: slides.
- Preserving statistical validity in adaptive data analysis. STOC 2015: slides.
- Approximate resilience, monotonicity, and the complexity of agnostic learning. SODA 2015: slides.
- Sample complexity bounds on differentially private learning via communication complexity. COLT 2014; ITA 2015: slides.
- Using data privacy for better adaptive predictions. Foundations of Learning Theory workshop @ COLT 2014 : slides.
- On the power and the limits of evolvability. Simons Institute workshop on Computational Theories of Evolution, 2014: slides.
- Optimal bounds on approximation of submodular and XOS functions by juntas. Simons Institute workshop on Real Analysis at @FOCS 2013 : slides.