Speaker: Miika Nikula
Title: Probabilistic Machine Learning for Probabilists

Abstract: 
We give an overview of probabilistic methods in machine learning 
with a special focus on inference for generative models. 
Probabilistic estimates are important for several reasons. 
In applications such as medicine it is important that any estimates 
are provided with well-calibrated uncertainty bounds. 
More intrinsically, probabilistic estimation provides a principled way 
of regularization and guarding against overfitting. The main challenge 
in formulating and obtaining probabilistic estimates in machine learning 
is the computational cost: while a point estimate of parameters can 
often be obtained in time O(N) for a dataset of N samples, 
a probabilistic estimate such as Gaussian process regression can be 
as costly as O(N^3). Various approximative methods have been developed 
to deal with the unfavorable scaling of time complexity, two of the 
most influential general-purpose methods being variational Bayesian 
inference and the family of stochastic gradient MCMC methods. 
We present an introduction to these methods and examples of their 
application in generative modeling.