One of the presentations will be by Anna Wigren, doctoral student supervised by LiU researcher Fredrik Lindsten, but located at Uppsala University. The work concerns how parts of machine learning can be automated in order to make it more efficient, more robust and easier to use.
“We must be able to manage uncertainty in the calculations to make the systems robust”, says Fredrik Lindsten, associate professor at the Division of Statistics and Machine Learning, Department of Computer and Information Science.
Computer models of complex systemsFredrik Lindsten’s research aims to develop and improve the algorithms used to build computer models of complex systems. Several different basic algorithms can be combined Fredrik Lindsten Photo credit Magnus Johanssonand modified, depending on the type of problem to be solved.
“One example in which there are many unknown parameters is calculating how epidemics spread. Our research has improved the algorithm for this”, says Fredrik Lindsten.
The researchers have related the result of the calculations to what has happened in the past, for example how diseases have spread around the world, how many people have contracted them, and where these people are located.
“We base our work on observations, and then use statistical methods to study how things change with time. This allows us to draw conclusions about what will happen in the future. The model is based on the assumption that if we know what happened before, we can predict the future”, says Fredrik Lindsten.
The same type of algorithm used to predict epidemics, for example, can also be used in completely different fields, such as oceanography. In this case, the way in which sea levels vary with time and under other influences is modelled in the system.
“We adapt the algorithm and the parameters to the data that have been collected”, he says.
Probabilistic programmingThey have integrated the method into something known as probabilistic programming, in which writing the code and the algorithms are independent of the specific details of the model to be used.
The term “model” here denotes the equations that are used to describe how, for example, an epidemic develops with time. Those who have insight into the application itself (in this case epidemiologists) are best qualified to specify the model. While constructing the model, knowledge about the application to be modelled is used.
On the other hand, it is seldom the epidemiologists, oceanographers or other users of probabilistic modelling who are the most capable of writing the code for the algorithm used to adapt the parameters in these models to the observations.
“The principle of probabilistic programming is that the user should be able to code the model in a generic programming language, after which the computational algorithms take Anna Wigrenover and adapt the models to the data that have been collected”, he says.
At the conference, Anna Wigren plans to use an example from an island in Micronesia showing the spread of dengue fever. This disease is spread by mosquito bites and kills
thousands of people every year.
Fredrik Lindsten and Anna Wigren have worked together with researchers at Uppsala University and the research department of the companyUber, Uber AI. The research has received funding from the Swedish Research Council and the Swedish Foundation for Strategic Research, and is part of the Wallenberg AI Autonomous Systems and Software Program, financed by the Knut and Alice Wallenberg Foundation.
Parameter elimination in particle Gibbs sampling, Anna Wigren, Riccardo Sven Risuleo, Lawrence Murray and Fredrik Lindsten, 33rd Conference on Neural Information Processing Systems (NeurIPS 2019), Vancouver, Canada.
Translated by George Farrants