COMP90051: Statistical and Evolutionary Learning

50% final exam (3 hours), 10% mid-semester test, 2x 20% projects.
The exam and total project marks are hurdles. If your mid-semester mark is lower than your average project mark, the mid-sem mark is dropped.

This subject covers a wide variety of techniques used in (choose your preferred buzzword) machine learning, data mining, Big Data, etc. Basically all of the methods of analysing data that don't fit under the traditional banner of "statistics". Because it covers so much, you don't go into very much depth in any of the topics in the lectures, but you do get a good idea of what methods are out there and what circumstances you might want to use them in. This subject is very much a case of "you get out what you put in", and sometimes feels a bit muddled when trying to explain mathematical ideas without using any actual maths, which is why I've only rated it 4/5.

The assessment during semester is in the form of open-ended projects which allow you to explore the methods in more detail and actually apply them to a practical task. The first one was about analysing social network data, trying to predict where users lived based on their friends and the time of day they were active. The second was handwriting recognition. Both projects were fun but challenging - expect to put in a lot of time if you want to do well. The first project had a competition website with a live leaderboard so you could see how well you were doing compared to the rest of the class. The second project was apparently supposed to too but the course coordinator didn't have time to set it up.

There are three lecturers for this course. Ben Rubinstein took the first half of the course in a "topic of the week" format, covering a lot of methods with little depth. Justin Bedo taught neural networks (3 weeks) and Vinh Nguyen taught evolutionary algorithms (3 weeks). Both Ben and Justin have experience working in the industry, Ben at Google and Justin at IBM. Of the three, Ben was my favourite lecturer, although I may be biased because I already knew him before taking the course... I least enjoyed the evolutionary algorithms part of the course, which could be summed up as "hey cool, this trick works in nature and it works when you implement it on a computer too". Other people might love it, though.

There is a prerequisite subject listed, a computer science subject on "Knowledge Technologies". In practice, the most important knowledge to have is programming experience (ideally in a high-level language suited for data analysis, e.g. Matlab, Python or R) and some probability and calculus. The lectures try to avoid going too deep into the maths, and there's an "intro to probability" document handed out at the start of the subject, but to get the most out the course, you'll need a little bit of maths.

The specific topics covered apparently vary a bit from year to year. This year we looked at:
- linear and logistic regression
- ensemble methods: bagging and boosting
- regularisation, model complexity and overfitting
- Support Vector Machines and kernel methods
- Probabilistic Graphical Models and Hidden Markov Models
- neural networks and "large scale learning" (methods for parallel computing etc)
- evolutionary/genetic algorithms for optimisation

Rambling aside: From the perspective of a mathematician, "machine learning" looks a whole lot like "statistics", but the focus is different. In statistics, the data you're dealing with usually has a nicely structured interpretation, and you want to answer specific questions within that framework. It's about understanding the real-world process that generated the data rather as much as it is answering questions about the data itself. In machine learning, the data is usually big, messy and unstructured, and all you care about is being able to make accurate predictions about future observations. Different approaches for different situations!

Yes, with screen capture.

Dr Ben Rubinstein, Dr Justin Bedo, Dr Vinh Nguyen.

No, but a practice exam was made available. The content of this subject has changed quite a bit in the last few years and is likely to be different again next year.

4/5

There are a couple of recommended texts. "The Elements of Statistical Learning" by Hastie, Tibshirani and Friedman covers most of the course and can be downloaded from the authors' web site.

2x one-hour lectures, 1x one-hour computer lab

2014, Semester 2.

H1