Advanced Machine Learning

Course#: CSCI-GA.3033-007

Instructor: Mehryar Mohri

Graders/TAs: Andrés Muñoz Medina.

Mailing List

Course Description

This course introduces and discusses advanced topics in machine learning. The objective is both to present some key topics not covered by basic graduate ML classes such as Foundations of Machine Learning, and to bring up advanced learning problems that can serve as an initiation to research or to the development of new techniques relevant to applications.

- Advanced standard scenario:
- Learning kernels.
- Deep ensemble methods.
- Structured prediction.

- On-line learning scenario:
- On-line learning basics.
- Learning and games.
- Learning with large expert spaces.
- Online convex optimization.
- Bandit problems.
- Sequential portfolio selection.

- Large-scale learning:
- Dimensionality reduction
- Low-rank approximation.
- Large-scale optimization.
- Distributed learning.
- Clustering.
- Spectral learning.
- Massive multi-class classification.

- Other non-standard learning scenarios:
- Domain adaptation and sample bias correction.
- Transduction and semi-supervised learning.
- Active learning.
- Time series prediction.
- Privacy-aware learning.

It is strongly recommended to those who can to also attend the Machine Learning Seminar.

Location and Time

Warren Weaver Hall Room 101,

251 Mercer Street.

Tuesdays 5:10 PM - 7:00 PM.

Prerequisite

Students are expected to be familiar with basic machine learning concepts and must have attended a graduate ML class such as Foundations of Machine Learning or equivalent, at Courant or elsewhere.

Projects and Assignments

There will be 2 homework assignments and a topic presentation and report. The final grade is a combination of the assignment grades and the project grades. The standard high level of integrity is expected from all students, as with all Math and CS courses.

Lectures

- Lecture 01: Learning kernels.
- Lecture 02: Deep ensemble methods.
- Lecture 03: Structured prediction.
- Lecture 04: Online learning basics.
- Lecture 05: Learning and games.
- Lecture 06: Learning with large expert spaces.
- Lecture 07: Online convex optimization.
- Lecture 08: Bandit Problems.
- Lecture 17: Domain adaptation.
- Lecture 18: Transduction.
- Lecture 19: Active Learning.

Technical Papers

An extensive list of recommended papers for further reading is provided in the lecture slides.

Homework

- Homework 1 [solution]
- Homework 2 [solution]
- Topic presentations:
- Learning with a Large Number of Experts: Component Hedge Algorithm,
Giulia DeSalvo and Vitaly Kuznetsov.

Reference: W. M. Koolen, M. K. Warmuth, and J. Kivinen. Hedging structured concepts. In COLT, pages 93–105, 2010. - Bandit Online Convex Optimization, Nadejda Drenska and Scott Yang.

Reference: Abraham Flaxman, Adam Tauman Kalai, H. Brendan McMahan. Online convex optimization in the bandit setting: gradient descent without a gradient. SODA 2005: 385-394. - Learning bounds for importance weighting, Tamás Madarász and Michael Rabadi.

Reference: Corinna Cortes, Yishay Mansour, and Mehryar Mohri. Learning bounds for importance weighting. In NIPS. 2010. - On the hardness of domain adaptation, Tamás Madarász.

Reference: Shai Ben-David and Ruth Urner. On the Hardness of Domain Adaptation and the Utility of Unlabeled Target Samples. In ALT. 2012 - Unlabeled Data: Now It
Helps, Now It Doesn’t, Mark Andrew Ward and Max Kuang.

Reference: A. Singh, R. D. Nowak, and X. Zhu. Unlabeled Data: Now It Helps, Now It Doesn’t. In NIPS, 2008. - Fast Global Alignment Kernels, Rodrigo Frassetto Nogueira and Thanos Papadopoulos.

Reference: Marco Cuturi. Fast Global Alignment Kernels. In ICML pp. Marco Cuturi: Fast Global Alignment Kernels. ICML 2011: 929-936, 2011. - A Spectral Algorithm for
Learning Hidden Markov Models, Kentaro Hanaki.

Reference: Daniel Hsu, Sham Kakade, and Tong Zhang. A Spectral Algorithm for Learning Hidden Markov Models. In COLT, 2009.

- Learning with a Large Number of Experts: Component Hedge Algorithm,
Giulia DeSalvo and Vitaly Kuznetsov.