CS 6961: Structured PredictionFall 2014

Course Information

2

Building up structured output prediction

• Refresher of binary and multiclass classification– A couple of weeks

• Simple structures– Show that multiclass is really a trivial kind of a structure– Training without inference

• Sequence labeling problems – HMM, inference, Conditional Random Fields, Structured variants of SVM and Perceptron

• Conditional models: How previous algorithms extend to general models• Complexity of inference and inference algorithms

• Different training regimes– Training with/without inference– Constraint driven learning, posterior regularization

• Learning without full supervision– Latent variables, semi-supervised learning, indirect supervision

• Advanced topics in inference

3

Class focus

• To see different examples– Sequence labeling, eg. Part-of-speech tagging– Predicting trees, eg Parsing– More complex structures, eg: relation extraction, object recognition, – And most importantly,

Your favorite domain/problem…

• To understand underlying concepts– Defining models, training, inference– Using domain knowledge to

• Define features• Define models• Make better predictions

4

The rest of this course

• The list may look complicated– You are right to complain– It is meant to be perplexing!

• Course objectives– To be able to define models, identify or design training and

inference algorithms for a new problem– To be able to critically read current literature

• We will revisit these slides in the last lecture!

5

Course mechanics

• Course structure– Lectures by me initially and gradually, presentations by you

• No text book– Some useful background reading on course website

• Machine learning is a pre-requisite

• Assignments (due dates on website/handout, more as we progress)– Three paper reviews (not hand written, please!)– One class presentation– One class project in groups of size 2– No midterm/final. Instead, project proposal, intermediate checkpoints, and final

report and presentationQuestions?

6

What assistance is available for you?

• Announcements on the course home pagehttp://svivek.com/teaching/structured-prediction/fall2014

• Lectures– Slides available on website– Additional notes and readings also on website

• Email: svivek@cs.utah.edu– Important: Prefix subject with CS6961

• Office hours:– MEB 3126– Tuesdays 2-3 PM, Thursdays 3:30-4:30 PM– Or by appointmentQuestions?

7

Policies (see website/handout for details)

• Collaboration vs. Cheating– Collaboration is strongly encouraged, cheating will not be tolerated– The School of Computing policy on academic misconduct

• If you haven’t already done this, read and sign the SoC policy acknowledgement form within two weeks

– Acknowledge sources and discussions in all deliverables

• Late policy – For reviews, a total of 48 hours of credit for the entire semester– Use these hours however you want– Not applicable to project

• Access– If you need any assistance, please contact me as soon as possible Questions?

8

Course expectations

This is an advanced course aimed at helping you navigate recent research.

I expect you to• Participate in the class

• Complete the readings for the lectures

• And most importantly, demonstrate independence and mathematical rigor in your work

9

• No readings for next lecture

• Please fill out and return the information sheet by the next lecture

• For questions about registration, please meet me now

Questions?