Deep probabilistic regression - new models and learning algorithms

Learning systems of the kind we develop in this project constitute core technology in all model-based engineered systems acting and interacting autonomously in a physical environment. We will develop and analyze new deep probabilistic regression models and learning algorithms. Modelling and learning nonlinear dynamics is given special attention.

Our project consists of three strategically coupled research themes, each with its own concrete ideas. 1) Develop probabilistic representations that make use of the flexibility offered by deep neural networks to create new formulations of regression problems.

Establish new theory to explain the overparameterized regime where deep learning models typically reside. 2) Establish and make use of the fundamental mechanisms allowing us to use sequential Monte Carlo within other learning algorithms. 3) Construct and learn flexible models of nonlinear dynamics, specifically deep energy-based models capable of representing nonlinear dynamical phenomena that generalize well to new data.This project opens up a new research direction compared to our ongoing VR project.

In particular, Theme 1 is expected to result in significant performance improvements in many applied areas in the same way as deep learning has already reshaped several areas via improved classification performance.

We have already established the necessary collaborations allowing us to pursue this project together with leading researchers around the world.