Machine Learning - Hilary Term 2016

About

Mathematics of machine learning. Overview of supervised, unsupervised and reinforcement learning. The course will cover neural networks, support vector machines, regression, clustering, PCA, collaborative filtering, as well important notions such as maximum likelihood, regularization and cross-validation.

Pre-requisites

Machine learning is a mathematical discipline; students will benefit from a solid background in probability, linear algebra and calculus. Programming experience and basic understanding of algorithms is essential. If you have doubts about your background, you should come talk to the lecturer as soon as possible.

(HTF) T. Hastie, R. Tibshirani and J. Friedman. The Elements of Statistical Learning. Springer 2011. (Available for download on the authors' web-page.)
(Mur) K. P. Murphy. Machine Learning: A Probabilistic Perspective, MIT Press 2012. (Electronic copy available through the Bodleian library.)
(Bis) C. M. Bishop. Pattern Recognition and Machine Learning. Springer 2007.
(Hay) S. Haykin. Neural Networks and Learning Machines. Pearson 2008.
(Nie) M. Nielsen. Neural Networks and Deep Learning. Online Book available here.
(GBC) I. Goodfellow, Y. Bengio, A. Courville. Deep Learning, MIT Press (in preparation). (Draft available here.)
(SB) R. S. Sutton and A. G. Barto. Reinforcement Learning: An Introduction. MIT Press (Online version and draft second edition here.)
(Sze) C. Szepesvári. Algorithms for Reinforcement Learning. Morgan and Claypool. (Online version here.)

Note: Copies of the first four textbooks are available for short loan in the department library. Further copies may also be available in the RSL and college libraries.

Examination

The course will have a take-home final exam which you will pick on monday of week 8 of Hilary term and hand in on the monday of week 0 of Trinity term. Note that MFoCS students have an earlier submission date as set by the Maths Institute.

Contact Information

Apart from lectures, classes and practicals, we have set up a Piazza forum. Please sign up here. You should use this forum to ask questions which could be answered by your classmates, TAs, and the lecturer. Should you so wish, you can also post to this forum anonymously. You may also send private messages to the lecturer and teaching staff, though we recommend that you use this option sparingly.

If your question is of a personal nature, you should contact the lecturer directly via email. Please allow for at least 48 hours for an email response.

In any class, it can be challenging for the lecturer to gauge how smoothly the class is progressing. We always welcome any feedback on what we could be doing better. If you would like to send anonymous comments or criticisms, please feel free to use an anonymous emailer like this one to avoid revealing your identity.

Lectures

Lecturer	Varun Kanade
Hours	16:00-17:00 on Wednesday and Friday
Location	Lecture Theatre B

Lecture Schedule

Date	Topic	Handouts	Reading
20/01/2016	Introduction to Machine Learning	[slides]	HTF Chap. 1 Mur Chap. 1 Bis Chap. 1 Hay Intro, Chap. 1
22/01/2016	Linear Regression	[before] [after]	HTF Chap. 2.1-3, Chap 3.1,3.2 Mur Chap. 7.1-2, 7.4 Bis Chap. 3.1
27/01/2016	Maximum Likelihood	[before] [after]	Normal distribution, Laplace distribution, Covariance, Entropy Mur Chap. 2.8, 4.1, 7.3-4 Bis Chap. 2.4, 3.2
29/01/2016	Basis Expansion, Regularization, Validation	[before] [after]	Mur Chap. 6.4-5, 7.1-2 HTF Chap. 3.4, 7 Bis Chap. 1.3-5, 3.1-2
03/02/2016	Regularization, Validation	[before] [after]	Mur Chap. 6.4-5, 7.1-2 HTF Chap. 3.4, 7 Bis Chap. 1.3-5, 3.1-2
05/02/2016	Optimisation	[before] [after]	Linear Programming, Gradient, Hessian, Gradient Descent, SGD Tricks (Léon Bottou)
10/02/2016	Optimisation, Logistic Regression	[before] [after]	Mur Chap 8
12/02/2016	Multi-class Logistic Regression, SVMs	[before][after]	Mur Chap 14.5 HTF Chap 4.5, Chap 12
17/02/2016	SVMs, Kernel Methods	[before][after]	Mur Chap 14 HTF Chap 4.5, Chap 12
19/02/2016	Neural Networks, Backpropagation	[before][after]	Nie Chap 1, 2 GBC Chap 6
24/02/2016	Training Neural Networks	[before][after]	Nie Chap 3, 5 GBC Chap 7,8
26/02/2016	Convolutional Networks	[before][after]	Nie Chap 6 GBC Chap 9
02/03/2016	Dimensionality Reduction and Multidimensional Scaling	[before][after]	HSF Chap 14.5, 14.8 Mur Chap 12.2-4
04/03/2016	Clustering	[before][after]	HSF Chap 14.3, 14.8 Mur Chap 25.4-5
09/03/2016	Clustering, Reinforcement Learning	[slides]	SB Chap 1, 3
11/03/2016	Reinforcement Learning	[slides]	SB Chap 1, 3

Classes

Classes will be held during weeks 3-7. All classes except Groups 7 and 9 will be in Room 051 Wolfson Building; Group 7 will be in Room 013 in the Robert Hooke Building; Group 9 will be in Room 114 in the Robert Hooke Building. You must hand in the problem sheets by noon on the Friday before the class.

Group 1	Monday	10:00-11:00	Abhishek Dasgupta
Group 2	Monday	11:00-12:00	Abhishek Dasgupta
Group 3	Wednesday	11:00-12:00	Rodrigo Mendoza Smith
Group 4	Wednesday	12:00-13:00	Rodrigo Mendoza Smith
Group 5	Friday	10:00-11:00	Francisco Marmolejo
Group 6	Friday	11:00-12:00	Francisco Marmolejo
Group 7^*	Monday	14:00-15:00	Syed (Ali) Rizvi
Group 8	Tuesday	16:00-17:00	Justin Bewsher
Group 9^†	Friday	11:00-12:00	Syed (Ali) Rizvi

^* This class will be in Room 013 in the Robert Hooke Building.

^† This class will be in Room 114 in the Robert Hooke Building.

Problem Sheets

Please hand in your sheets before noon on Friday the week before the class.

Problem Sheet 1 (Week 3)
Problem Sheet 2 (Week 4)
Problem Sheet 3 (Week 5)
Problem Sheet 4 (Week 6)
Problem Sheet 5 (Week 7)
Problem Sheet 6 (Week 8) (not for submission; no classes in week 8)

Practicals

Practicals will be held in weeks 2-8 in Room 379. The demonstrators for practicals are Abhishek Dasgupta, Bernardo Pérez Orozco, Marija Marcan.

Group 1	Tuesday	10:00-12:00
Group 2	Thursday	09:00-11:00
Group 3	Thursday	11:00-13:00

Practicals will use Torch, a powerful programming framework for machine learning, in particular deep learning, that is very popular at Google and Facebook research.

Installation: The script provided in practical 1 will install torch on lab machines. For those using torch on their personal machines, we are unable to provide any support. The installation process is quite easy though and there is plenty of help available online.

All the practicals will be available on the Github repository.

Week 2	Learning Lua and the tensor library	pdf
Week 3	Linear Regression	pdf
Weeks 4, 5	Optimization and MNIST Classification	pdf
Week 6	MNIST Classification using SVM	pdf
Week 7	Implementing a new layer	pdf
Week 8	Experimenting with PCA	pdf

Lecturer

Varun Kanade
Office Hours: Tuesday 15:30-16:30 in Room 449 Wolfson Building

Classes

Groups	Instructor
1, 2	Abhishek Dasgupta
3, 4	Rodrigo Mendoza Smith
5, 6	Francisco Marmolejo
7, 9	Syed (Ali) Rizvi
8	Justin Bewsher

Practicals

Marija Marcan

ICML Test of Time Award Papers

2015	Learning to Rank using Gradient Descent C. Burges, T. Shaked, E. Renshaw, A. Lazier, M. Deeds, N. Hamilton, G. Hullender
2014	Multiple Kernel Learning, Conic Duality, and the SMO Algorithm F. Bach, G. Lanckriet, M. Jordan
2013	Semi-Supervised Learning Using Gaussian Fields and Harmonic Functions X. Zhu, Z. Ghahramani, J. Lafferty Online Convex Programming and Generalized Infinitesimal Gradient Ascent M. Zinkevich
2012	Diffusion Kernels on Graphs and Other Discrete Input Spaces R. Kondor, J. Lafferty
2011	Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data J. Lafferty, A. McCallum, F. Pereira
2010	Reducing Multiclass to Binary: A Unifying Approach for Margin Classifiers E. Allwein, R. Schapire, Y. Singer
2009	Transductive Inference for Text Classification using Support Vector Machines T. Joachims Policy invariance under reward transformations: Theory and application to reward shaping A. Ng, D. Harada, S. Russell Least-Squares Temporal Difference Learning J. Boyan
2008	Combining Labeled and Unlabeled Data with Co-Training A. Blum, T. Mitchell