Probabilistic machine learning
Principal lecturer: Dr Damon Wischik
Taken by: MPhil ACS, Part III
Code: L173
Term: Michaelmas
Hours: 16 (8 x 2hrs lectures)
Class limit: max. 12 students
Prerequisites: For Cambridge students, a background in machine learning and Bayesian inference is required, and knowledge of deep neural networks is recommended. External students should have a similar background in probability and machine learning.
timetable
Aim
Teach the probabilistic basis of advanced techniques in neural network modelling
Syllabus
Deep learning as "probabilistic modelling with ultra-high dimensional models". It covers:
(1) supervised learning and generative modelling as maximum likelihood fitting,
(2) variational, adversarial, and energy-based bounds for fitting distributions,
(3) approaches to generalization / inference, in particular, approaches to modelling uncertainty,
(4) causal deep learning,
(5) some standard models / modalities, including transformers and their link with Gaussian processes.
The course will describe this material in the style of "applied probability modelling", but it will also emphasize the practical aspect of running these models on actual real datasets.
Objectives
Equip students to understand and apply advanced theory in deep learning
Assessment
- Three take-home mini essay / practicals (20% each). These will involve some programming and investigation, and a write up of about 2 pages.
- mini-project worth 40%, which will ask them to apply techniques from a recent paper to a novel dataset or task and analyse the results.