/MathsDL-spring19

Mathematics of Deep Learning, Courant Insititute, Spring 19

MathsDL-spring19

Topics course Mathematics of Deep Learning, NYU, Spring 19. CSCI-GA 3033.

Logistics

  • Mondays from 7.10pm-9pm. CIWW 102

  • Tutoring Session with Parallel Curricula (optional): Fridays 11am-12:15pm CIWW 101.

  • Piazza: sign-up here

  • Office Hours: Tuesdays 4:30pm-6:00pm, office 612, 60 5th ave.

Instructors

Lecture Instructor: Joan Bruna (bruna@cims.nyu.edu)

Tutor (Parallel Curricula): Luca Venturi (lv800@nyu.edu)

Tutor (Parallel Curricula): Aaron Zweig (az831@nyu.edu)

Syllabus

This Graduate-level topics course aims at offering a glimpse into the emerging mathematical questions around Deep Learning. In particular, we will focus on the different geometrical aspects surounding these models, from input geometric stability priors to the geometry of optimization, generalisation and learning. We will cover both the background and the current open problems.

Besides the lectures, we will also run a parallel curricula (optional), following the Depth First Learning methodology. We will start with an inverse curriculum on the Neural ODE paper by Chen et al.

Detailed Syllabus

  • Introduction: the Curse of Dimensionality

  • Part I: Geometry of Data

    • Euclidean Geometry: transportation metrics, CNNs , scattering.
    • Non-Euclidean Geometry: Graph Neural Networks.
    • Unsupervised Learning under Geometric Priors (Implicit vs explicit models, microcanonical, transportation metrics).
    • Applications and Open Problems: adversarial examples, graph inference, inverse problems.
  • Part II: Geometry of Optimization and Generalization

    • Stochastic Optimization (Robbins & Munro, Convergence of SGD)
    • Stochastic Differential Equations (Fokker-Plank, Gradient Flow, Langevin Dynamics, links with SGD; open problems)
    • Dynamics of Neural Network Optimization (Mean Field Models using Optimal Transport, Kernel Methods)
    • Landscape of Deep Learning Optimization (Tensor/Matrix factorization, Deep Nets; open problems).
    • Generalization in Deep Learning.
  • Part III (time permitting): Open qustions on Reinforcement Learning

Pre-requisites

Multivariate Calculus, Linear Algebra, Probability and Statistics at solid undergraduate level.

Notions of Harmonic Analysis, Differential Geometry and Stochastic Calculus are nice-to-have, but not essential.

Grading

The course will be graded with a final project -- consisting in an in-depth survey of a topic related to the syllabus, plus a participation grade. The detailed abstract of the project will be graded at the mid-term.

Final Project is due May 1st by email to the instructors

Lectures

Week Lecture Date Topic References
1 1/28 Guest Lecture: Arthur Szlam (Facebook) References
2 2/4 Lec2 Euclidean Geometric Stability. Slides References
3 2/11 Guest Lecture: Leon Bottou (Facebook/NYU) Slides References
4 2/18 Lec3 Scattering Transforms and CNNs Slides References
5 2/25 Lec4 Non-Euclidean Geometric Stability. Gromov-Hausdorff distances. Graph Neural Nets Slides References
6 3/4 Lec5 Graph Neural Network Applications Slides References
7 3/11 Lec6 Unsupervised Learning under Geometric Priors. Implicit vs Explicit models. Optimal Transport models. Microcanonical Models. Open Problems Slides References
8 3/18 Spring Break References
9 3/25 Lec7 Discrete vs Continuous Time Optimization. The Convex Case. Slides References
10 4/1 Lec8 Discrete vs Continuous Time Optimization. Stochastic and Non-convex case Slides References
11 4/8 Lec9 Gradient Descent on Non-convex Optimization. Slides References
12 4/15 Lec10 Gradient Descent on Non-convex Optimization. Escaping Saddle Points efficiently. Slides References
13 4/22 Lec11 Landscape of Deep Learning Optimization. Spin Glasses, Kac-Rice, RKHS, Topology. Slides References
14 4/29 Lec12 Guest Lecture: Behnam Neyshabur (IAS/NYU): Generalization in Deep Learning Slides References
15 5/6 Lec13 Stability. Open Problems. References

Lab sessions / Parallel Curricula

DistributionalRL: Living document

  • Class 1: Basics of RL and Q learning
    • Required Reading:
      • Sutton and Barto (Ch 3, Ch 4, Ch 5, Ch 6.5)
        • The standard introduction to RL. Focus in Chapter 3 on getting used to the notation we’ll use throughout the module, and an introduction to the Bellman operator and fixed point equations. In Chapter 4 the most important idea is value iteration (and exercise 4.10 will ask you to show why iterating the Q function is basically the same algorithm).
        • Chapter 5 considers using full rollouts to estimate our value / Q function, rather than the DP updates. Focus on the difference between on-policy and off-policy, which will be relevant to the final algorithm.
        • Including 6.5 is an introduction to Q-learning in practice, updating one state-action pair at a time (without worrying about function approximation yet).
      • Contraction Mapping Theorem (3.1)
        • We’ll need the notion of contractions repeatedly throughout the module. Their essential property is a unique fixed point, and you should have a clear understanding of the constructive proof of this fixed point (don’t worry about the ODE applications).
    • Questions:
      • Exercise 3.14, Exercise 4.10 in S & B
      • Prove the Bellman operator contracts Q functions with regard to the infinity norm
      • What is a sanity-check lower bound on complexity for Q learning? Why might this be infeasible for RL problems in the wild?

NeuralODE: Living document