Probabilistic Machine Learning
Pre-reading responses (5% × 4)
Format. Half a page pdf each. During the term, you will be asked to talk through at least one of your responses, either during a lecture or in a brief individual viva after the lecture. LLM usage is permitted.
Write a short response answering the question below. Most of the questions are about judgement, and ask you to read between the lines or think about the paper's unstated premise, or explain an argument that is implied rather than directly stated. You get the marks for interpretation and engaging critically with the paper, not for simply repeating what it says.
Be concrete. For example, point to a specific passage, equation, or figure, say why it matters and what it does. Use the course's notation, not the paper's. Answer from the paper and the lectures only; facts brought in from elsewhere earn no marks, however true: the question is what this paper says and how it relates to this course. Engaging critically means setting the paper against the lectures, not against the later literature.
- Session 2:
Read Goodfellow et al. §§1–4
and
submit on Moodle
by 19 Oct 4pm.
(i) Proposition 1 gives the optimal discriminator D*G(x). Write down its logit, log D* / (1 − D*). The proof assumes the discriminator sees equally many real and generated samples; say where that assumption enters. (ii) State equation (5) in the course's notation, q for the truth and pθ for our model, for the variant in which the discriminator sees real and generated samples in proportion π to 1 − π. (iii) In lecture 1, our target was KL(q ‖ pθ). What arguments does the paper give in favour of using (5) rather than our target? Comment. - Session 4.
TBC - Session 6.
TBC - Session 8.
TBC
Exercise 1 (15%) and Exercise 2 (15%)
TBC
Proposal (5%) and Written Report (45%)
TBC