Invited talks

The Invited Talks at IDEAL 2024 bring together experts to share their latest research and practical insights straightforwardly and engagingly. These sessions are an excellent opportunity to explore new ideas, tackle current challenges, and see how cutting-edge techniques are applied in the real world. This year, we are very proud to host two invited talks: 

Inverse Problems in the Age of AI 

Inverse problems involve reconstructing unknown physical quantities from indirect measurements. They appear in various fields, including medical imaging (e.g., MRI, Ultrasound, CT), material sciences and molecular biology (e.g., electron microscopy), just to name a few examples. Deep neural networks are currently able to achieve state-of-the-art performance in many imaging tasks and in this talk we argue that  in inverse imaging problems efficient deep neural networks with more predictable performances can only be achieved by combining model-based solvers with learned models. There is plenty of evidence for this and typical examples where this integration has had an impact include the plug-and-play framework and the network unfolding strategy. 

In the first part of the talk, we introduce INDigo+, a novel INN-guided probabilistic diffusion algorithm for arbitrary image restoration tasks. INDigo+ combines the perfect reconstruction property of invertible neural networks (INNs) with the strong generative capabilities of pre-trained diffusion models. Specifically, we leverage the invertibility of the networks to condition the diffusion process and in this way we generate high quality restored images consistent with the measurements. 
 
In the second part of the talk, we discuss the unfolding techniques which is an approach that allows embedding priors and models in the neural network architecture. In this context we discuss the problem of monitoring the dynamics of large populations of neurons over a large area of the brain. Light-field microscopy (LFM), a type of scanless microscopy, is a particularly attractive candidate for high-speed three-dimensional (3D) imaging which is needed for monitoring neural activity. We review fundamental aspects of LFM and then present computational methods for neuron localization and activity estimation from light-field data. Our approach is based on machine learning and leverages the intrinsic characteristics of neuronal signals and the physics of the acquisition process. 

Finally, we look at the multi-modal case and present an application in art investigation. Often X-ray images of Old Master paintings contain information of the visible painting and of concealed sub-surface design, we therefore introduce a model-based neural network capable of separating from the “mixed X-ray”  the X-ray image of the visible painting and the X-ray of the concealed design.  

 
This is joint work with  A. Foust, P. Song, C. Howe, H. Verinaz, J. Huang, Di You and Y. Su from Imperial College London, M. Rodrigues and W. Pu from University College London, I. Daubechies from Duke University and C. Higgitt and N. Daly from The National Gallery in London.

---

Bringing Deep Learning and Reasoning closer

Deep Learning continues to attract the attention and interest not only of the wider scientific community and industry, but also society and policy makers. However, the mainstream approach of end-to-end iterative training of a hyper-parametric, cumbersome, and opaque model architecture led some authors to brand it “black box”. Cases were reported when such models can give wrong predictions with high confidence – something that jeopardises the safety and trust. Deep Learning is focused on accuracy and overlooks explainability and the semantic meaning of the internal model representations, reasoning and its link with the problem domain. In fact, it shortcuts from the large amount of (labelled) data to the predictions bypassing and substituting the causality with correlation and error minimisation. It relies on assumptions about the data distributions that are often not satisfied and suffers from catastrophic forgetting when faced with continual and open set learning. Once trained, such models are inflexible to new knowledge. They are good only for what they were originally trained for. Indeed, the ability to detect unseen and unexpected and start learning this new class/es in real time with no or very little supervision (zero- or few- shot learning) is critically important but is still an open problem. The challenge is to fill the gap between the high levels of accuracy and the semantically meaningful solutions.

This talk will focus on “getting the best from both worlds”: the powerful latent feature spaces formed by pre-trained deep architectures such as transformers combined with the interpretable-by-design (in linguistic, visual, semantic, and similarity-based form) models. One can see this as a fully interpretable frontend and a powerful backend working in harmony. Examples will be demonstrated from the latest projects from the area of autonomous driving, Earth Observation, health and a set of well-known benchmarks.