Machine learning, signal processing, and image analysis

Theme DSB1. Visual Intelligence: Beyond applying Machine Learning to Images

• Contact person: Adin Ramirez Rivera
• Keywords: Computer vision, Image processing, Machine Learning, Probabilistic Models, Visual Intelligence, Histopathology
• Research group: Digital Signal Processing and Image Analysis (DSB)

This research theme focuses on advancing machine learning (ML) methods for analyzing complex image data, particularly within the domains of computer vision and image processing. The goal is to use and innovate upon diverse methodologies, with an emphasis on probabilistic models and transformer technologies, to enhance the analysis and interpretation capabilities of ML techniques when confronted with intricate image datasets including those with thousands of high-resolution gigapixel images. The research will involve both the application of existing ML strategies to complicated image contexts and the active development of novel approaches to push the boundaries of what is currently possible. In addition to the DSB group, the candidate will be collaborating closely with the Visual Intelligence Center as well as engaging in interdisciplinary work combining ML and cancer research in the field of computational pathology with the Oslo University Hospital. This broad collaboration will enrich the research with insights from various expert groups, aiming to yield significant contributions to the applied fields as well as to ML in general.

Methodological research topics:

• Probabilistic Machine Learning
• Self-supervised learning
• Interpretable and explainable models
• Models incorporating physical, anatomical, or geometrical constraints
• Models including uncertainty estimation
• Leveraging high-resolution gigapixel images

Possible Projects:

• Developing Visual Intelligent applications with our partners (contact Anne H Schistad Solberg)
• Principled Probabilistic Machine Learning: Navigating Unsupervised Representations (contact Ad¨ªn Ram¨ªrez Rivera)
• Understanding Flexible Tokenization on Transformers (contact Ad¨ªn Ram¨ªrez Rivera)
• New approaches for training more accurate and robust deep learning models in computational pathology (contact Andreas Kleppe)

External Partners:

• We have close collaboration with Norwegian companies like GE Healthcare, the Institute of Marine Research, Kongsberg Satellite Services, and Equinor, as well as the university hospitals, in particular the Institute for Cancer Genetics and Informatics at Oslo University Hospital whom we have been collaborating with on ML topics for 30 years using huge datasets (multiple petabytes of data). Access to real-world datasets and challenging real problems give large opportunities for new research of significant impact to the domains.

Theme DSB3. Understanding Memory and Emotion with Machine Learning

• Contact persons: Ali Ramezani-Kebrya and Ad¨ªn Ram¨ªrez Rivera
• Keywords: Machine Learning, Human Memory and Emotion, Time Series, Anomaly Detection, Complex Data
• Research group: Digital Signal Processing and Image Analysis (DSB)

Understanding human cognition is a highly impactful topic that encompasses Neuroscience, Psychology, and Computer Science. We use ML to gain deeper insights into human Memory and Emotion.

Re Memory, our collaborative team is equipped with unique neural activity measurements from epilepsy patients, which allows us not only to decipher visual processing and memory but also to uncover the causes of sensory and memory deficits in pathological brain states. Our ultimate ambition is to revolutionize the diagnosis of memory disorders by establishing new biomarkers and refining therapeutic approaches.

Re Emotion, we build on our preliminary results to train Machine Learning models to recognize chronic and severe stress and fatigue, and extend to recognize emotions across various types of personalities using multiple data types, including images, videos, and physiological signals captured as time-series data. Our research has the potential to significantly enhance both diagnostic and therapeutic practices and ultimately improve human well-being.

Successful candidates should be able to navigate the multidisciplinary intersection of neuroscience, machine learning, psychology, and signal processing.

Potential Projects:

• Understanding Memory (Contact Ad¨ªn Ram¨ªrez Rivera, J?rgen Afseth Sugar): Our collaborative team is equipped with unparalleled brain activity measurements alongside SotA algorithms in neurophysiology, machine learning, and signal processing. We have access to unique neural activity data from epilepsy patients. Beyond deciphering visual processing and memory, we uncover the causes of sensory and memory deficits in pathological brain states. Our ultimate ambition is to revolutionize the way memory disorders are diagnosed, to establish new biomarkers, and to refine therapeutic approaches. Specific methodologies include: time series encoding and processing, attention mechanism, anomaly detection, and multi-model learning.
• Understanding Emotion (Contact Ali Ramezani-Kebrya): In terms of technical expertise, this project involves anomaly detection, multi-model learning, and attention mechanism (transformer architectures). In terms of supervision, this project is supervised by Prof. Ali Ramezani-Kebrya (actively publishing in top ML venues, PI in Norwegian Centre of Excellence Integreat and SFI Visual Intelligence, ELLIS member, NeurIPS Area Chair) and Prof. Carsten Griwodz (Section Head of Digital Infrastructure and Security with several EU Funded projects) in collaboration with Institute for Energy Technology - IFE.

Potential partner: 

• Institute for Energy Technology - IFE, which we have an ongoing collaboration that is committed to advising us on the technology and developing impact cases (we are open to working with any relevant partner)

Theme DSB5. Graph Neural Networks for Hierarchical Structured Learning

• Contact person: Ad¨ªn Ram¨ªrez Rivera
• Keywords: Graphs, Graph Neural Networks, Machine Learning, Hierarchical Learning, Hierarchy Discovery
• Research group: Digital Signal Processing and Image Analysis (DSB)/Integreat

This research theme seeks to advance the capabilities of graph neural networks (GNNs) to pioneer a new frontier in hierarchical structure learning. By tapping into attention, flexible tokenization, self-supervised learning, among other innovations in ML, we aim to extract the rich structural information embedded within data, an endeavor that enables the distillation of knowledge into representations that act locally as well as within the global context of the data points. This research aims to design and learn networks with the ability to adapt to varying contexts, allowing for a more nuanced understanding of relational patterns. By augmenting GNNs with the dynamic range and context sensitivity, we anticipate pushing the boundaries of what these architectures can learn about hierarchical structures in data.

Possible methodologies:

• Graph-based tokenization for transformers inputs
• Self-supervised graph-learning for structure and graph embeddings at different levels
• Relation predictions based on Graph Neural Networks
• Hierarchical discovery through Graph Neural Networks
• Other relevant methodologies

Applications:

• We intend to validate the robustness and versatility of our models across different domains. Whether it's deciphering complex biological networks, illuminating social media dynamics, or optimizing transportation systems, the foundational research in graph-based hierarchical learning will set the stage for a suite of diverse downstream tasks. This cohesive and expansive research theme not only aligns with the scientific quest to understand increasingly complex data, but it also sets a new precedent for practical applications that stand to benefit from these enhanced machine learning capabilities.