Probabilistic Programming and its Semantics
About the speaker
Carol is a second year DPhil student at Oxford under the supervision of Prof Luke Ong. She is currently working on the semantics of probabilistic PL and deep probabilistic PL.
The scientific approach to obtain knowledge is to first observe a phenomenon in the world, make hypothesis and then to run enough experiments to convince ourselves that some conclusion can be drawn from the experimental data. Machine learning intends to mimic this scientific approach in a computer. One popular approach is Bayesian inference, where an estimated distribution of a random variable (prior distribution) is transformed, according to the likelihood of an observed data, to an updated distribution (posterior distribution). Intuitively, this updated distribution is the “knowledge” the computer gained by observing this particular data. Probabilistic programming language is a language that facilitates programmers to specify probabilistic models in a language such that their executions correspond to Bayesian inference. Many practical probabilistic PLs are equipped with posterior inference algorithms, and these languages are favoured by data scientists since it frees them from the design of inference algorithms and enable them to focus their time on the design of probabilistic models.