How DID YOU GET INTO MAchine LEARNING?
My postgraduate studies at the Department of Mathematical Statistics at the University of the Free State mostly focused on Bayesian statistics – both research and course work. While doing my masters, I moved to Pretoria and was allowed to replace one course module with a relevant module at the University of Pretoria. I chose a module on Artificial Intelligence presented by Prof Andries Engelbrecht. This was in 2000. I moved away from research for a while and started my PhD in 2007. The topic of my PhD thesis was on probabilistic topic modelling: an unsupervised pattern recognition technique which clusters documents into topics. I have now been working in this area for 10 years with special focus on unsupervised modelling of natural language text. Mainly this entails the extraction of patterns from large text corpora without prior knowledge or annotation (text mining). Of special interest is the application of topic modelling and other text mining techniques to short texts and understanding the modelling challenges posed by short texts. Some of my other research projects include wildlife protection and green security games.
WhAT WILL YOU Be teaching?
I will be speaking about unsupervised learning. Modelling techniques will include mixture models and latent variable models. Density estimation relies on EM and other related learning algorithms that I will cover. I will start with latent variable models for continuous data and move on to discrete data. At the end of the lecture you will know the difference between mixture models, PCA, probabilistic PCA, Factor Analysis to mention a few. In short – we will lay out the landscape of latent variable models. Evaluation metrics designed for supervised learning don’t hold for unsupervised learning and I will briefly discuss evaluation methods for unsupervised learning. I will point you to resources to get started in this field. You can expect the lecture to be in a Jupyter notebook format.
What TOPICs Excite you most at the moment?
Two topics that excite me most currently:
- Bayesian deep learning
- Word representation algorithms such as word2vec – and combining it with document representation models such as LDA.