In this talk we use Oura data examples to provide context why to measure sleep, how to measure sleep, how machine learning is used to turn the measurements into knowledge, and what to measure during sleep. In addition, some highlights will be showed on how data science can provide novel information when turned into population level analysis.

Join us to discover how AWS Sports customers leverage Generative AI and cutting-edge Large Language Models (LLMs) to create dynamic, real-time and engaging match commentary that captivates viewers. By seamlessly integrating live match data feeds and audio, the AI solution generates a vivid, constantly evolving narrative that immerses fans in the on-field action. The solution goes a step further by generating commentaries in multiple languages and writing styles to cater to diverse audiences. This innovative solution offers soccer enthusiasts a more engaging and thrilling viewing experience, further enhancing their love for the sport.

Climate Change is one of the defining issues of our time. Yet, navigating the vast amount of information - often complex and constantly evolving - can be overwhelming for the public, policymakers, and even scientists themselves. Large Language Models (LLMs) have the potential to be a game-changer, providing improved information retrieval, summarisation of research, and even support for decision-making. However, we must acknowledge the challenges that come with this power. Climate data can rapidly become outdated, regional variations demand nuanced approaches, and the increasing sophistication of LLMs can mask errors, making human supervision critical. This keynote will explore the opportunities and obstacles of utilising LLMs to combat climate change misinformation and empower informed action. We will delve into how LLMs can serve the public, policymakers, and scientists alike while acknowledging the need for responsible development and human-in-the-loop approaches to ensure reliable and equitable solutions.

Over the past few decades, machine learning has proven its value in improving many processes along the weather forecasting value chain: from obtaining observations, quality control and assimilation into forecast models, to improving forecasts and producing customised products. It is only in the last two years that deep learning has demonstrated its potential to replace the entire workflow. In this talk, I will present various machine learning applications that we are using and developing at MeteoSwiss, some of which have been operational for several years, and discuss the next challenges.

Machine learning has been an enabler in re/insurance over the last few decades with many success stories from automated underwriting engines and property underwriting using satellite imagery to fraud detection and image recognition for automotive claims processing. However, the ever-growing data landscape demands new approaches. Generative AI offers a transformative leap, unlocking the possibilities beyond the capabilities of traditional machine learning models. Thanks to the latest advancements in fine-tuning LLMs efficiently and tool-equipped agents, true game-changer solutions in the insurance sector are around the corner, even with the sector's factual accuracy and deep domain knowledge requirements.

This talk explores the potential of Generative AI in the insurance industry by delving into current example use cases from Swiss Re on specialized underwriting AI assistants and streamlined claims processing. It will address key challenges like explainability, data privacy, and regulatory requirements with the learnings from the use cases so far to enable more efficient processes and customer-centric user experience.

As vision plays a key role in how we interpret a situation, developing vision-based perception for robots promises to be a big step towards robotic navigation and intelligence, with a tremendous impact on automating robot navigation. This talk will discuss our recent progress in this area at the Vision for Robotics Lab of ETH Zurich and the University of Cyprus (http://www.v4rl.com), and some of the biggest challenges we are faced with.

Anomaly detection and localisation are critical for automated visual inspection. Current approaches are sensitive to domain shifts, can not handle high-resolution images and severely underperform in a low data regime. At IBM, we have developed an end-to-end framework based on masked multi-scale reconstruction for anomaly detection enhanced with visual prompting and automatic thresholding. Tested on real manufacturing and civil engineering scenarios, we are able to detect the smallest defects with only a few dozen images to train on. the pipeline will be released in IBM’s Maximo Application Suite and Watsonx in Q2’24.

Multi-Label Classification (MLC) is a common task in the legal domain, where more than one label may be assigned to a legal document. A wide range of methods can be applied, ranging from traditional ML approaches over fine-tuned Transformer-based architectures to LLM prompting. In this talk, we discuss an evaluation of different MLC methods using two public legal datasets, POSTURE50K and EURLEX57K. By varying the amount of training data and the number of labels, we observe the comparative advantage offered by different approaches in relation to the dataset properties. Our findings help kickstart MLC experimentation in new projects and highlight performance-speed-cost trade-off considerations.