End of course
The assessment
This page explains what the assessment examines and what to have ready. The paper itself, the data and the submission link appear here when the window opens.
What is examined
The assessment follows the same discipline as the four teaching days. It is not a test of recall: the vocabulary of deep learning is easy to repeat and earns little on its own. What is marked is whether you can take a claim from the data to a conclusion you can defend, and state what that conclusion does not cover.
- reading a model description and stating the shape of every object in it, and the number of parameters that follows from those shapes;
- separating a training result from held-out evidence, and naming which split chose the procedure;
- judging a reported number: which population it describes, what would make it stop applying, and whether the comparison behind it was fair;
- identifying the failure mode a given result is vulnerable to, such as a shortcut, a leaked split, a stale document, or an unverified action;
- stating plainly what a piece of evidence does not establish.
What to prepare
Set your machine up during the teaching days rather than on the morning of the assessment.
- a working Python installation with Jupyter, and the packages listed in the practical archive, installed and tested by running one of the day notebooks end to end;
- or a free Google Colab account, which needs no installation and runs in the browser;
- the practical archive unpacked and the Day 1 and Day 2 data prepared, since Day 2 downloads its source the first time it runs;
- a stable connection for downloading the materials and uploading your answer.
The practical archive and the data →
On using AI tools
You may use AI assistants, as you would at work. They produce code that runs and confident prose about results. They are much weaker at judging whether a number means anything, and that judgement is what is being marked. A submission that reproduces an assistant's output without checking it is usually easy to recognise.
Where you use an assistant, say so in the submission. That is expected and costs no marks.
