NLDL 2023 Winter School

Preliminary Program

Monday January 9th:

09:00: Registration

09:15: Opening

09:30: A gentle introduction to Deep Reinforcement Learning (Rudolf Mester and Even Klemsdal, NTNU Trondheim)

Reinforcement Learning has emerged as a powerful technique in modern machine learning, allowing a system to learn through trial and error. We will introduce some fundamental principles upon which this family of methods is based. In the first part, we give a summary of classical reinforcement learning and thus provide the essentials for understanding Deep Reinforcement Learning (DRL). In the second part, we shift the focus onto DRL and look at how deep learning enhances classical reinforcement learning, leading to powerful new algorithms. These algorithms include DQN playing Atari games at a superhuman level, PPO solving robot locomotion problems, and Alpha Zero learning to play GO, Chess, and Shogi. In the limited time of such a mini-tutorial, we can of course only sketch the main characteristics, but we intend to motivate to look deeper into this fascinating family of methods.

11:00: Self-Supervised Learning: Training Targets and Loss Functions (Zheng-Hua Tan, Aalborg University)

Humans learn much under supervision but even more without. Such will apply to machines. Self-supervised learning is paving the way by leveraging unlabeled data which is vastly available. In this emerging learning paradigm, deep representation models are trained by supervised learning with supervisory signals (i.e., training targets) derived automatically from unlabeled data itself. It is abundantly clear that such a learnt representation can be useful for a broad spectrum of downstream tasks with the need of significantly less supervised data.

As in supervised learning, key considerations in devising self-supervised learning methods include training targets and loss functions. The difference is that training targets for self-supervised learning are not pre-defined and greatly dependent on the choice of pretext tasks. This leads to a variety of novel training targets and their corresponding loss functions. This tutorial aims to provide an overview of training targets and loss functions developed in the domains of speech, vision and text. Further, we will discuss some open questions, e.g., transferability, and under-explored problems, e.g., learning across modalities. For example, the pretext tasks, and thus the training targets, can be drastically distant from the downstream tasks. This raises the questions like how transferrable the learnt representations are, how to choose training targets and how to explore representations. We will also discuss the link between semi-supervised and self-supervised learning.

12:15: Lunch

13:30: The Challenge of Unverfiability in eXplainable AI Evaluation (Anna Hedstrøm, TU Berlin)
In recent years, the interest in eXplainable AI (XAI) techniques has undoubtedly exploded and countless methods and toolboxes are now at the disposal of XAI researchers, ML practitioners and data scientists alike. Despite my shared enthusiasm for this productive development, until recently, the topic of XAI evaluation was grossly understudied — and caused confusion about which explanation methods work and under what conditions. In this tutorial, we will take an in-depth look at some of the most recent developments in XAI evaluation and review the solutions that the community has put forward to circumvent the fundamental problem of XAI — the lack of ground-truth explanations. In addition to this, we will discuss The Challenge of Unverfiability — why evaluation is so difficult to get right (and what we can do about it). At the end of the tutorial, a brief yet hands-on introduction to Quantus will be given, which is an open-source library intended for XAI researchers to evaluate local neural network explanations.

15:00: Tutorial 4 (Line Clemmensen, Technical University of Denmark)

16:30: Early career presentations/posters + mingling & finger food

January 10-12th: Main Conference Program

Friday January 13th:

09:00: Tutorial 6 (NRIS): High Performance Computing for Deep Learning Part 1

The tutorial will contain the following topics:

• Transition from a Laptop to a central server to a compute cluster.

• The SLURM queue manager.

• Software access on a shared resource.

• Example job scripts.

• Understanding resources.

10:40: Tutorial 7 (NRIS): High Performance Computing for Deep Learning Part 2

The tutorial will contain the following topics:

• Introducing the Deep Learning example that will provide the context for this tutorial.

• How to access GPUs.

• Data staging.

• Submit the analysis as a HPC job.

• Monitor job progress, debug errors and Interpret results.

• Summary.

12:00: Lunch

13:00: Topic: Representation learning and learning with few data (Marcus Liwicki, Luleå University of Technology)

Deep Neural Networks are data hungry, they require millions of labelled data in order to work! --- Really? --- The last decade has shown useful approaches to work with less labelled data, either by having a lot of data from a similar domain or by letting the network learn meaningful representations without explicit supervision. On Monday, you learned already about self-supervised learning. This tutorial first brings it in to general perspective of learning with few data, covering typical transfer learning and auto-encoder approaches or perceptual loss. Furthermore, the tutorial will investigate some typical (mis-) conceptions of these methods and suggests some practical tips on how to learn with few data. By participating in this tutorial, you will get deep insights in representation learning and learning with few data, as well as practical tools to start working on data in your own domain.