Current activities within the Arena
Learning Gaussian process models
Gaussian processes are flexible models that can be used to predict a wide range of real-world processes, from the spatial distribution of electromagnetic fields to the temporal variation of carbon dioxide levels. Such models are specified by a mean and covariance structure, which depend on the application, and learned from data. Importantly, Gaussian process models also provide a measure of uncertainty of the resulting predictions.
Machine learning in the study of the first galaxies
In a collaboration between CIM and astronomers at the Department of Physics and Astronomy, machine learning takes the stage as a powerful tool in the interpretation of data on the first generations of galaxies.
A machine-learning approach to measuring the escape of ionizing radiation from galaxies in the reionization epoch
Contact: Kristiaan Pelckmans, Erik Zackrisson (UU/Astro), Christian Binggeli (UU/Astro), Ruben Cubo.
Learning nonlinear dynamical systems using Sequential Monte Carlo
Sequential Monte Carlo (SMC) methods allows us to approximate a sequence of probability distributions defined on a sequence of spaces of increasing dimension. One important case where this is highly useful is then it comes to the problem of learning the states and parameters in a nonlinear dynamical system. This is the classic application of SMC and it is often referred to as the particle filter in this setting. For this problem we develop new methods, associated analysis and importantly we also work with relevant real world industrial applications.
Sequential Monte Carlo methods for system identification
Particle Gibbs with Ancestor Sampling
Linking social behavior to the brain
Which part of the brain is related to social behavior? This project blends experimental results based on behavior of groups of guppies (Poecilia Reticulata) with advanced neuroimaging techniques for answering this question.
Contact: Kristiaan Pelckmans, David Sumpter (UU/Math), Niclas Kolm (SU/Bio), Hongli Zeng (UU/Math), Alex Sorokovsky (UU/Math).
Feature-Augmented Deep Neural Networks for Segmentation of Cells
We use a fully convolutional neural network for microscopy cell image segmentation. Rather than designing the network from scratch, we modify an existing network to suit our dataset. We demonstrate improved cell segmentation, modality transfer learning and the ability to segment irregularly shaped cells. The method is demonstrated on datasets consisting of phase contrast, fluorescent and bright-field images.
Deep neural network architecture for E. Coli cell segmentation
Network science and engineering
Networks are models to describe systems of interconnected entities, such as individuals and their social relations, airports and their connections, or proteins and their interactions. At the Uppsala University InfoLab we work on the analysis of networks, specializing in social and online information networks. We apply existing machine learning methods to study networks as a service for researchers and companies, and we develop new methods for complex network data with various types of attributes. Our research results include algorithms for dimensionality reduction, cluster analysis and positional analysis.
Contact: Matteo Magnani
Inference in graphical models using Sequential Monte Carlo methods
Performing inference in a general graphical model involves for example finding the joint distribution of the latent variables, finding unknown parameters in the model and computing the so-called partitioning function (the normalization constant). The latter object is of interest for example in information theoretical problems of channel capacity computations and in statistical mechanics when it comes to computing the free energy of a system. These inference problems are hard in the sense that there are no closed form expressions (except for in simple special cases). We work on deriving new computational solutions based on Sequential Monte Carlo (SMC) and the associated analysis.
Deep Learning–Based Classification of Zebrafish Deformation for High-Throughput Screening
Zebrafish is an important vertebrate model organism in biomedical research, especially suitable for morphological screening due to its transparent body during early development. Here we demonstrate the potential of a deep learning approach for accurate high-throughput classification of whole-body zebrafish deformations in multifish microwell plates. The input is raw image data, obviating the need of expert knowledge for feature design or optimization of the segmentation parameters.
Deformation probability for the raw input images, images with the tails of the fish alone and images with the heads of the fish alone for the treated and untreated fish samples.