Type: Seminar

Distribution: World

Expiry: 8 Oct 2018

CalTitle1: Introduction to Singular Learning Theory

Auth: erich@10.83.64.53 (ehes5653) in SMS-WASM

Hello all, The next MaPSS talk of this semester will be at 17:00 on Mon 13th August in Carslaw 535. It’s a great opportunity to meet fellow postgrads, listen to an interesting talk, and of course get free food! ************************************************************************************** Speaker: Haruki Osaka Title: Introduction to Singular Learning Theory Abstract: Many statistical models used in data science have hierarchical structures and hidden variables, for example, mixture models, hidden Markov models, neural networks and so on. Although such models are widely used in practice, no sound theoretical foundation for the large sample behaviour of these models has been established. The main reasons for this difficulty is that of non-identifiability and a degenerate Fisher Information metric of these models, which are basic regularity conditions required for Fisher’s asymptotic normal theory. Such statistical models are called singular. Statistical inference for singular models are different to that of regular models. In addition, popular information criteria used in model selection such as AIC and BIC are not valid for singular models. Singular Learning Theory is a new area of mathematics which attempts to use algebraic techniques to study singular models. Using these results, Drton and Plummer (2017) recently proposed the singular Bayesian information criterion (sBIC) that is valid for singular models. In this talk, I will give an overview of some typical problems that may occur in singular models and how singular learning theory attempts to resolve them. ************************************************************************************** See you there! Details can also be found on the school’s new Postgraduate Society website: http://www.maths.usyd.edu.au/u/MaPS/mapss.html Cheers, Eric