Course Description

Deep learning is a rapidly growing segment of machine learning. It is increasingly used to deliver near-human level accuracy for many tasks such as image understanding, multimodal reasoning, voice recognition, natural language processing, and agentic automation. Applications span medical imaging, drug discovery, scientific research, robotics and autonomous vehicles, as well as personalised education, financial modelling, cybersecurity, climate forecasting, and industrial automation.

This 5-day course is built upon tutorials on specific topics on deep learning. The course further encompasses among others keynote talks, as well as special sessions on industry and diversity in AI as part of the NLDL conference program. 

The Winter School will provide a study of several emerging topics of high relevance within advanced deep learning, from a basic understanding of the techniques to the latest state-of-the-art developments in the field. In addition, the participants will be exposed to the latest advances and applications in deep learning by the oral presentations and poster presentations in the main conference program.

The course will thus consist of 5 full days of the NLDL conference including: tutorials, keynote sessions, oral presentations and poster presentations, as well as practical components.

The Winter School offers 2.5 ECTS credits. Upon successful completion of the course, participants will receive a certificate confirming their completion of the Winter School. This certificate may be used to register the credits at their home universities. 

Winter School Tutorial and Speakers

The NLDL Winter School will consist of tutorials from leading experts in the machine learning field. Below we present each tutorial, followed by the tutorial speakers. 

Learning with Order: Deep Ordinal Representations for Trustworthy AI

Many imaging tasks involve labels that are inherently ordinal, (e.g., such as disease severity grades and cancer stages in medical imaging). These labels exhibit a clear order but lack fixed distances between categories. Despite this, most AI pipelines treat ordinal labels as either nominal (using cross-entropy loss) or continuous (using regression losses such as MSE). This mismatch leads to brittle models, implausible errors, and topologically inconsistent predictions—particularly in safety-critical settings.

This tutorial presents Deep Ordinal Learning as a unifying inductive bias for robust and trustworthy AI. Rather than viewing ordinal methods as task-specific fixes, we provide a coherent, end-to-end perspective on how ordinal structure can be explicitly incorporated across the modelling pipeline.

The tutorial begins with a brief introduction to the motivation and structure of the session, followed by four core technical sections:

1. Robust Ordinal Classification: rank-consistent architectures, unimodal constraints, and ordinal-aware losses that explicitly enforce order and reduce clinically severe misclassifications, especially in small, noisy, and imbalanced datasets.

2. Ordinal Segmentation: extension of ordinality to dense prediction tasks using pixel-wise ordinal losses to enforce topological and anatomical consistency in nested or layered structures.

3. Closing the Metric Gap: ordinal-specific evaluation protocols (e.g., Quadratic Weighted Kappa, Uniform Ordinal Classification Index) that better reflect clinical risk than nominal or regression metrics.

4. The Frontier – Ordinal Representations and Generative AI: ordinal constraints in representation learning and generative models

A hands-on component using PyTorch will reinforce practical adoption.

Learning objectives:

1. understand when and why ordinal modelling is appropriate

2. learn how to design ordinal-aware classifiers and segmentation models

3. apply appropriate ordinal evaluation metrics

4. gain insight into emerging ordinal approaches for representation learning and generative AI.

The tutorial will be presented by Valentina Corbetta from the The Netherlands Cancer Institute and Wilson Silva from Utrecht University.

Evaluating Large Language Models: Beyond Accuracy and Human Evaluation

How do we know whether a Large Language Model (LLM) is actually good for a specific task? This tutorial will focus on LLM evaluation, it will firstly discuss why evaluating LLMs is fundamentally different from evaluating classifiers or other machine learning models, it will introduce traditional Natural Language Processing (NLP) evaluation metrics, such as BLEU and ROUGE, as well as human evaluation approaches and discuss the issues with these methods when applied to LLMs. The tutorial will then introduce a new-ish paradigm in LLM evaluation, called LLM-as-a-judge, provide an introduction to the current state of the art, discuss potential limitations such as biases and touch on issues such as traceability and attribution. The tutorial will finish with a hands-on exercise where participants will have the chance to try some of these models

The tutorial will be presented by Dimitra Gkatzia from Edinburg Napier University

Arctic AI: Earth System Modeling and Evaluation 

Recent advances in machine learning have made AI-based weather forecasting increasingly competitive with traditional numerical weather prediction. The Arctic is a particularly important region for these developments: it plays a key role in the global climate system and is undergoing rapid environmental change, with far-reaching implications for ecosystems, local communities, and geopolitics. At the same time, sparse observations, complex terrain, and strongly coupled atmosphere–ocean–sea ice processes make reliable forecasting particularly challenging, motivating the development of specialized machine learning approaches for Arctic applications.

This tutorial provides an overview of AI-based weather forecasting, with a focus on specialized applications such as sea ice prediction and high-resolution wind forecasting in Norwegian fjords. We will then explore ensemble forecasting, uncertainty quantification, and metrics for evaluating predictions. These topics will be further explored in a hands-on exercise in which participants work with an ensemble of sea ice concentration forecasts and investigate how different metrics capture forecast quality and uncertainty.

The tutorial will be presented by Lars Uebbing from UiT The Arctic University of Norway and Andrew MacDonald from the University of Cambridge. 

Agentic AI and AI-Assisted Research

This tutorial will give machine learning researchers a hands-on, realistic roadmap for using agentic AI to accelerate their scientific work. We will start with the core building blocks: how models, context, prompts, and execution harnesses fit together into functioning systems. Moving beyond browser chat, we will work directly with local CLI tools and APIs, showing how to use modern "vibe-coding" workflows to rapidly build, test, and iterate on functional code directly on your machine.

The second half will focus on AI-assisted research in practice. We will explore how to equip agents with custom skills, define clear goals, and structure closed execution loops. We will see how to scale from single tasks to coordinating sub-agents and multi-agent teams. Finally, we will look at how to organize and run AI-conducted experiments: handling automated runs and baseline explorations inspired by "The AI Scientist" paper.

The tutorial will be presented by Benjamin Ricaud and Daniel Kaiser from UiT The Arctic University of Norway.

Contact : 

Winter School Chair:

• Puneet Sharma (puneet.sharma@uit.no)

• Kristoffer Wickstrøm (kristoffer.k.wickstrom@uit.no)

• Elisabeth Wetzer (elisabeth.wetzer@uit.no)

For other inquiries: 

• Robert Jenssen (robert.jenssen@uit.no)

• Michael Kampffmeyer (michael.c.kampffmeyer@uit.no).