IFT 6169
Theoretical principles for deep learning
Winter 2023 · hybrid, in English
- Lectures
- Wednesday and Thursday, 9h30-11h15
- First class
- Thursday, January 12th, 2023
- Location
- Mila’s Auditorium 2, 6650 Saint-Urbain, first floor (no key card needed); also online
- Office hours
- Thursday 11:15am-12:00pm, right after class
- Instructor
- Ioannis Mitliagkas
- Platforms
- StudiUM (announcements), Slack (discussion), Gradescope (quizzes and reports)
Description
Research in deep learning produces state-of-the-art results on a number of machine learning tasks. Most of those advances are driven by intuition and massive exploration through trial and error. As a result, theory is currently lagging behind practice. The ML community does not fully understand why the best methods work.
- Why can we reliably optimize non-convex objectives?
- How expressive are our architectures, in terms of the hypothesis class they describe?
- Why do some of our most complex models generalize to unseen examples when we use datasets orders of magnitude smaller than what the classic statistical learning theory deems sufficient?
A symptom of this lack of understanding is that deep learning methods largely lack guarantees and interpretability, two necessary properties for mission-critical applications. More importantly, a solid theoretical foundation can aid the design of a new generation of efficient methods—sans the need for blind trial-and-error-based exploration.
In this class we will go over a number of recent publications that attempt to shed light onto these questions. Before discussing the new results in each paper we will first introduce the necessary fundamental tools from optimization, statistics, information theory and statistical mechanics. The purpose of this class is to get students engaged with new research in the area. To that end, the majority of credit will be given for a class project report and presentation on a relevant topic.
Tentative topics
- Generalization: theoretical analysis and practical bounds
- Information theory and its applications in ML (information bottleneck, lower bounds etc.)
- Generative models beyond the pretty pictures: a tool for traversing the data manifold, projections, completion, substitutions etc.
- Taming adversarial objectives: Wasserstein GANs, regularization approaches and controlling the dynamics
- The expressive power of deep networks (deep information propagation, mean-field analysis of random networks etc.)
Prerequisites
This is a graduate class for students who want to engage in theory-driven deep learning research. We will introduce some theoretical tools necessary.
- Probability
- Linear algebra
- Machine learning
Class format and communication
The class is hybrid:
- In person at Mila’s Auditorium 2 (located at 6650 Saint-Urbain, first floor. No key card needed to access)
- Online via Zoom
- Video recordings of lectures will be made available after lectures end.
We will use the class’s Studium to make announcements. We will use a Slack workspace for discussion. Please keep an eye out for an invitation to the Slack. We will use Gradescope for quizzes and report submissions.
If you have something personal/sensitive to discuss, feel free to email me or the TA. Starting your email subject with ‘IFT6169:’ will ensure that your email is not miscategorized.
Registering and auditing
If you want to audit the class, please fill this form. You will get access to the course material, but you will not be able to submit work or be graded.
Evaluation
| Component | Weight |
|---|---|
| Class project on a research topic | 35% |
| Paper presentations | 25% |
| Quizzes | 20% |
| Midterm exam | 15% |
| Class participation | 5% |
Use this LaTeX template for scribing.
Schedule and materials
Introduction
-
Jan 12
Crash course in optimization
For the first half of the class we will be closely following the previous iteration of the class.
-
Jan 18Jan 19
Basics of convex analysis and gradient descent
Reading
Convex analysis basics from ‘Convex Optimization’ by Boyd, Vandenberghe ([5] under references):
- Chapter 2 (required: beginning of chapter to 2.1.4, recommended: 2.1.5 to end of section)
- Chapter 3 (required: beginning of chapter to 3.1, recommended: 3.2, 3.3 and 3.4)
Convergence proofs: from Chapter 3 of [1] (‘Convex Optimization…’ by S.Bubeck under References)
- Required: Convergence proof of Theorem 3.2 (note that we studied the unconstrained case. In this case the projection operator PiX(x) is the identity operator)
-
Jan 25
The different rates of gradient descent: from Lipschitz to strongly convex
Reading
Convergence proofs from Chapter 3 of [1] (‘Convex Optimization…’ by S.Bubeck under References)
- Required: Sections 3.1, 3.2, 3.4, with emphasis on the Convergence proof of Theorem 3.12
- Optional: Section 3.3 and anything related to constrained optimization. Keep in mind that we still discuss in class simplified unconstrained versions of some constrained results from Bubeck. Anything in our scribed notes is also required.
-
Jan 26
Black box models and lower bounds
-
Feb 1
Accelerated methods: Polyak's momentum (the heavy ball method)
-
Feb 2
Nesterov's Accelerated Gradient, Stochastic gradient descent
Crash course in statistical learning theory
-
Feb 8
Stochastic gradient descent
-
Feb 9Feb 15
Elements of statistical learning theory
-
Feb 16
PAC-Bayes bounds
-
Feb 22
Probably Approximately Correct Constrained Learning
-
Feb 23
Midterm exam
-
Mar 1Mar 2
Winter break, no class
-
Mar 8
Stability and generalization
Seminar part of class
-
Mar 9
Weighted Sums of Random Kitchen Sinks: Replacing minimization with randomization in learning
Reading
- Recommended reading: Section 16 (Kernel methods) of [4]
- Required reading: [23]
- Recommended viewing: video of talk by A. Rahimi. The discussion pertaining to this paper, starts at 31:10 (though the intro might be necessary) and ends at about 48:10.
-
Mar 15
Surprising results on the generalization error of neural networks
Reading
- Recommended reading: In Search of the Real Inductive Bias: On the Role of Implicit Regularization in Deep Learning
- Recommended reading: Understanding deep learning requires rethinking generalization
- Recommended reading: On the Bias-Variance Tradeoff: Textbooks Need an Update, blog post by Brady Neal
-
Mar 16
Where is the Bayes risk hiding?
Reading
- Recommended reading (refresher from last week): Understanding deep learning requires rethinking generalization
- Recommended reading: To Understand Deep Learning We Need to Understand Kernel Learning
- Recommended reading: Overfitting or perfect fitting? Risk bounds for classification and regression rules that interpolate
-
Mar 22Mar 23Mar 29Mar 30Apr 5
Student presentations
-
Apr 13
Neural Tangent Kernel (new lecture, no scribed notes)
Reading
- Recommended reading: Neural Tangent Kernel: Convergence and Generalization in Neural Networks
- Recommended reading: Blog summary of paper
Recommended reading on reproducing kernel hilbert spaces:
- Guest lecture by Jose Gallego-Posada from last year: video
- [15min] Watch this great video on duality.
- [5min] Get familiar with the statement of the Riesz representation theorem and think about the relationship with the video above.
- [40min] Read the Wikipedia entry on RKHSs - Can skip sections 4, 8 and 9.
- [15min] Optional: Read sections 1 and 2 of the Wikipedia entry on Kernel embedding of distributions
-
Apr 19Apr 20
1-1 discussions on course projects
-
May 3
Final poster presentation for projects
Extra material (not scheduled)
-
Determinantal point processes for ML (guest lecture by Jose Gallego-Posada)
Reading
- Recommended reading: Notes from Stefanie Jegelka’s class at MIT, written by Swati Gupta.
- Recommended reading: Determinantal Point Processes in Randomized Numerical Linear Algebra, Derezinski and Mahoney.
Bonus material from past iterations
Older lectures that later material replaced. Not examined.
-
Dynamical systems, with applications to neural networks (guest lecture by Guillaume Lajoie)
-
Distributional reinforcement learning
-
Expressivity and universal approximation
Reading
Michael Nielsen’s visual proof; Section 20 to 20.4 of [4]; [40]
-
Basic results on reinforcement learning
References
Numbers in the readings above refer to this list.
- Convex Optimization: Algorithms and Complexity, Sebastien Bubeck.
- Theory of classification: a survey of some recent advances Stephane Boucheron, Olivier Bousquet and Gabor Lugosi
- iPython notebook demonstrating basic ideas of gradient descent and stochastic gradient descent, simple and complex models as well as generalization.
- Understanding Machine Learning: From Theory to Algorithms, by Shai Shalev-Shwartz and Shai Ben-David.
- Convex Optimization, Stephen Boyd and Lieven Vandenberghe.
- Nesterov’s Accelerated Gradient Descent for Smooth and Strongly Convex Optimization, blog post by Sebastien Bubeck.
- Introductory lectures on convex optimization, Yurii Nesterov.
- Why momentum really works, blog post by Gabriel Goh (this blog post uses a slightly different parametrization of the momentum algorithm. The version we discuss in class, only applies the learning rate on the gradient.)
- YellowFin and the Art of Momentum Tuning, SysML 2019, J. Zhang, I. Mitliagkas.
- Large-scale Machine Learning and Optimization (class), Dimitris Papailiopoulos, University of Wisconsin.
- Advanced Machine Learning Systems (class), Chris De Sa, Cornell University.
- A PAC-Bayesian Tutorial with A Dropout Bound, David McAllester.
- Stability and generalization, O. Bousquet, A. Elisseeff.
- Train faster, generalize better: Stability of stochastic gradient descent, M. Hardt, B. Recht, Y. Singer.
- Computing Nonvacuous Generalization Bounds for Deep (Stochastic) Neural Networks with Many More Parameters than Training Data, Gintare Karolina Dziugaite, Daniel M. Roy
- Lecture notes on generative learning algorithms, Andrew Ng
- Generative Adversarial Nets, Ian Goodfellow et al.
- Wasserstein GAN, Martin Arjovsky, Soumith Chintala, Léon Bottou
- Read-through: Wasserstein GAN, Alex Irpan
- The Numerics of GANs, Lars Mescheder, Sebastian Nowozin, Andreas Geiger
- Optimization Methods for Large-Scale Machine Learning, Léon Bottou, Frank E. Curtis, Jorge Nocedal
- Accelerating Stochastic Gradient Descent using Predictive Variance Reduction, Rie Johnson, Tong Zhang
- Weighted Sums of Random Kitchen Sinks: Replacing minimization with randomization in learning, Ali Rahimi, Ben Recht
- PacGAN: The power of two samples in generative adversarial networks, Zinan Lin, Ashish Khetan, Giulia Fanti, Sewoong Oh
- Stochastic First- and Zeroth-order Methods for Nonconvex Stochastic Programming, Saeed Ghadimi, Guanghui Lan
- Linear Convergence of Gradient and Proximal-Gradient Methods Under the Polyak-Łojasiewicz Condition, Hamed Karimi, Julie Nutini, Mark Schmidt
- Escaping From Saddle Points — Online Stochastic Gradient for Tensor Decomposition, Rong Ge, Furong Huang, Chi Jin, Yang Yuan
- How to Escape Saddle Points Efficiently, Chi Jin, Rong Ge, Praneeth Netrapalli, Sham M. Kakade, Michael I. Jordan
- RANDOM MATRICES AND COMPLEXITY OF SPIN GLASSES, ANTONIO AUFFINGER, GERARD BEN AROUS, AND JIRI CERNY
- The Loss Surfaces of Multilayer Networks, Anna Choromanska, Mikael Henaff, Michael Mathieu, Gérard Ben Arous, Yann LeCun
- Learning Long-Term Dependencies with Gradient Descent is Difficult, Y Bengio, P Simard, P Frasconi
- On the difficulty of training Recurrent Neural Networks, Razvan Pascanu, Tomas Mikolov, Yoshua Bengio
- Opening the Black Box: Low-Dimensional Dynamics in High-Dimensional Recurrent Neural Networks, David Sussillo and Omri Barak
- Reinforcement Learning: An Introduction, Richard S. Sutton and Andrew G. Barto
- Course Notes for EE227C (Spring 2018): Convex Optimization and Approximation, Moritz Hardt.
- Uniform convergence may be unable to explain generalization in deep learning, Vaishnavh Nagarajan, J. Zico Kolter
- Tutorial on Practical Prediction Theory for Classification, John Langford, 2005.
- Computing Nonvacuous Generalization Bounds for Deep (Stochastic) Neural Networks with Many More Parameters than Training Data, Gintare Karolina Dziugaite, Daniel M. Roy, 2017.
- Non-Vacuous Generalization Bounds at the ImageNet Scale: A PAC-Bayesian Compression Approach, Wenda Zhou, Victor Veitch, Morgane Austern, Ryan P. Adams, Peter Orbanz, 2018.
- Representation benefits of deep feedforward, Matus Telgarsky, 2015.
Class bibliography
You can find a large number of recent (mostly) research papers related to the class’s objectives here. You can use those for your in-class paper presentations and projects.