Deep Structured Learning (IST, Fall 2018)

Summary

Structured prediction is a framework in machine learning which deals with structured and highly interdependent output variables, with applications in natural language processing, computer vision, computational biology, and signal processing. In the last 5 years, several applications in these areas achieved new breakthroughs by replacing the traditional feature-based linear models by more powerful deep learning models based on neural networks, capable of learning internal representations.

In this course, I will describe methods, models, and algorithms for structured prediction, ranging from "shallow" linear models (hidden Markov models, conditional random fields, structured support vector machines) to modern deep learning models (convolutional networks, recurrent neural networks, attention mechanisms, etc.), passing through shallow and deep methods akin to reinforcement learning. Representation learning will also be discussed (PCA, auto-encoders, and various deep generative models). The theoretical concepts taught in this course will be complemented by a strong practical component, by letting students work in group projects where they can solve practical problems by using software suitable for deep learning (e.g., Pytorch, TensorFlow, DyNet).

Course Information

Instructor: André Martins
TAs: Vlad Niculae, Erick Fonseca
Schedule: Wednesdays 14:30-18:00, Room LT2 North Tower Level 4 (tentative)
Communication: http://piazza.com/tecnico.ulisboa.pt/fall2018/pdeecdsl

Grading

Homework assignments (60%)
Final project (40%)

Project Examples

The course project is an opportunity for you to explore an interesting problem using a real-world dataset. You can either choose one of our suggested projects or pick your own topic (the latter is encouraged). We encourage you to discuss your project with TAs/instructors to get feedback on your ideas.

Team: Projects can be done by a team of 2-4 students. You may use Piazza to find potential team mates.

Milestones: There are 3 deliverables:

Proposal: A 1-page description of the project. Do not forget to include a title, the team members, and a short description of the problem, methodology, data, and evaluation metrics. Due on 17/10.
Midway report: Introduction, related work, details of the proposed method, and preliminary results if available (4-5 pages). Due on 14/11.
Final report: A full report written as a conference paper, including all the above in full detail, finished experiments and results, conclusion and future work (8 pages excluding references). Due on 12/12.

All reports should be in NIPS format. There will be a class presentation and (tentatively) a poster session, where you can present your work to the peers, instructors, and other community members who will stop by.

See here for a list of project ideas.

Recommended Bibliography

Deep Learning. Ian Goodfellow and Yoshua Bengio and Aaron Courville. MIT Press, 2016.
Machine Learning: a Probabilistic Perspective. Kevin P. Murphy. MIT Press, 2013.
Linguistic Structured Prediction. Noah A. Smith. Morgan & Claypool Synthesis Lectures on Human Language Technologies. 2011.

Schedule

Date	Topic	Optional Reading
Sep 19	Introduction and Course Description	Goodfellow et al. Ch. 1-5 Murphy Ch. 1-2
Sep 26	Linear Classifiers	Murphy Ch. 3, 6, 8-9, 14	HW1 is out!
Oct 3	Feedforward Neural Networks	Goodfellow et al. Ch. 6
Oct 10	Neural Network Toolkits Guest lecture: Erick Fonseca	Goodfellow et al. Ch. 7-8	HW1 is due. HW2 is out!
Oct 17	Linear Sequence Models	Smith, Ch. 3-4 Murphy Ch. 17, 19	Project proposal is due.
Oct 24	Representation Learning and Convolutional Neural Networks	Goodfellow et al. Ch. 9, 14-15
Oct 31 (rescheduled to Oct 29, rooms V1.17/V1.11!)	Structured Prediction and Graphical Models	Murphy Ch. 10, 19-22 Goodfellow et al. Ch. 16 David MacKay's book, Ch. 16, 25-26	HW2 is due. HW3 is out!
Nov 7	Recurrent Neural Networks	Goodfellow et al. Ch. 10
Nov 14 (room E5)	Sequence-to-Sequence Learning	Sutskever et al., Bahdanau et al., Vaswani et al.
Nov 21 (room F8)	Attention Mechanisms and Neural Memories Guest lecture: Vlad Niculae	Learning with Sparse Latent Structure	HW3 is due. HW4 is out!
Nov 28	Deep Reinforcement Learning Game of Taxi Guest lecture: Francisco Melo		Midterm report is due.
Dec 5	Deep Generative Models	Goodfellow et al. Ch. 20 Murphy, Ch. 28 NIPS16 tutorial on GANs Kingma and Welling, 2014
Jan 9			Final report is due.
Jan 16	Final Projects I
Jan 23	Final Projects II