New technique boosts accuracy and efficiency in probabilistic programming
Programming can be a time-consuming and error-prone task. KTH researchers propose an automated technique that relieves developers of manual processes.
KTH doctoral students Daniel Lundén, Gizem Çaylak and Professor David Broman, together with Fredrik Ronquist from the Swedish Museum of Natural History in Stockholm, have identified a way to significantly improve the efficiency and accuracy of statistical inference applied in probabilistic programming languages (PPLs).
PPLs allow users to encode statistical problems and automatically apply algorithms to solve them. Statistical inference refers to the uncertainty surrounding conclusions drawn from a limited sample of empirical data. This uncertainty frequently appears in a broad range of research fields.
Automatic optimisation
Lundén and his colleagues have developed an automatic technique that, for any given probabilistic program, determines checkpoints in probabilistic programs that always run in the same order. This alignment technique will relieve developers of manually identifying checkpoint locations.
“Alignment is a property in PPLs that researchers have discussed for a long time,” Lundén explains.
“Our work is unique because we develop a technique that automatically aligns probabilistic programs. In previous work, PPL users must manually align programs, which is tedious and error-prone.”
By automating the process of determining checkpoint locations, researchers and practitioners can focus on the core aspects of dealing with statistical inference problems without the need for manual optimisation.
Relevant for a broad range of research
The alignment technique is particularly important for phylogenetics—the study of evolutionary history among groups of organisms such as animal species. Specifically, the alignment technique is critically important when determining how often species evolve into new species and how often species go extinct.
However, the new technique can also benefit probabilistic programming in other research fields, such as computer vision, topic modelling and cognitive science.
The paper won the EAPLS Best Paper Award at the European joint conferences on theory and practice of software (ETAPS) 2023.
Read the paper: Automatic Alignment in Higher-Order Probabilistic Programming Languages
Contact
Text: Sturle Hauge Simonsen
Related news
Students design computer games in class
Do you want to learn how to develop your own computer game? Or are you curious about what the market strategy looks like in the computer game industry? Maybe you'd rather create graphics? Then this co...
Read the article
New technique boosts accuracy and efficiency in probabilistic programming
Programming can be a time-consuming and error-prone task. KTH researchers propose an automated technique that relieves developers of manual processes.
Read the article