Total suggested work for this week

4 hours lectures (including first neural network lecture), 2 hours computer lab session, 5-10 hours on mini-project. Total 11-16 hours.

Before you start...

We prepared a self-assessment quiz with true-false statements from previous exams about important concepts of week 4. The quiz does not count towards the final grades and can be repeated multiple times. Note, however, that it does not cover everything that you should know about the topics of week 4.

5.1 Mini-project

The primary focus of your time this week should be the mini-project. You now have all the tools you need to be able to produce a very good answer to this problem. Please ask questions during help-desk sessions.

5.2 Feature selection: lecture

Monday 23rd November 10:15 - 12:00 we will have a lecture on feature selection and applications.

Lecture notes are available here.

5.3 Feature selection: lab session

In this lab on Wednesday 25th (15:15-17:00) we will analyse two data sets (happiness comparison between countries and goals in football).

The updated computer labs are available here:

Note: Make sure to download the updated version (24 November 13:00) that uses scikit-learn instead of statsmodels!

Answers available here.

These will be covered at the tutorial on Wednesday 25th November 15:15-17:00. Please attempt the questions before the tutorial. Also feel free to ask questions in the discussion section.

5.4 The artificial intelligence debate

Are humans or AI better problem solvers? In the lecture we look at how using our own intelligence (doing feature selection ourselves) we can produce models that give insight in to the systems we study. I also talk about how data scientists work in practice. But there are problems which machine learning outperforms humans. Most notably, when playing board games.

Neural network methods can also perform well on ar arcade games where most game information is seen on the screen. For example, pong and space invaders.

But other games, even quite simple ones like Ms Pacman, cannot be solved without 'telling' the neural network which features are important.

This has led to a debate between AI researchers about how far we can come with machine learning without combining it with some other form of reasoning. You can hear Yann LeCunn (Facebook AI) and Gary Marcus debate this issue here.

You can read Gary Marcus critique of deep learning here.