Mats Granath

Docent in Theoretical Physics @ University of Gothenburg

Quantum Machine Learning vs. Machine Learning for Quantum Computing

Machine learning and quantum computing are two expanding technologies with tremendous potential. What if they are combined, into quantum machine learning (QML)? I will give an overview of QML, pointing out that quantum mechanics makes it difficult to “quantize” classical ML algorithms such as artificial neural networks. Instead, existing QML algorithms are typically hybrid algorithms, part classical, part quantum. An alternative to making ML quantum is to use ML as part of the software required to operate a quantum computer. I will give some examples from my own research of both types of algorithms, using a hybrid QML approach to address a logistics problem and using deep reinforcement learning for compiling quantum code and for quantum error correction.

Olle Häggström

Professor of Mathematical Statistics @ Chalmers University of Technology

AI Alignment and Our Momentous Imperative to Get It Right

We are standing at the hinge of history, where actions taken today can lead towards a long and brilliant future for humanity, or to our extinction. Foremost among the rapidly developing technologies that we need to get right is AI. Already in 1951, Alan Turing warned that “once the machine thinking method had started, it would not take long to outstrip our feeble powers”, and that “at some stage therefore we should have to expect the machines to take control.” If and when that happens, our future hinges on what these machines’ goals and incentives are, and in particular whether these are compatible with and give sufficient priority to human flourishing. The still small but rapidly growing research area of AI Alignment aims at solving the momentous task of making sure that the first AIs with power to transform our world have goals that in this sense are aligned with ours.

John Kumar

Principal AI Platform Engineer @ AstraZeneca

How AstraZeneca Scientists, Data Quality and AI IT Teams Came Together To Accelerate the COVID Vaccine

In late 2020, against the backdrop of lockdowns and a global pandemic, AstraZeneca was striving to deliver a vaccine at cost. As the vaccine trials were drawing to a close, there was a need to accelerate the analysis of free-text reporting by doctors as part of the trial. There was a need to stand up both the infrastructure and a robust NLP model in a highly regulated, high stakes environment. Ultimately, the team managed to transform the COVID-19 vaccine reporting process from days to hours. In addition to laying the foundation for an enterprise machine learning platform, this work was a small part of releasing the vaccine, which has saved an estimated million lives, prevented 50 million infections, and delivered 2 billion doses.

Louise Anderson-Conway

AI Resident @ Google

Trajectory Ensembling

Traditionally, only the last checkpoint of a training run is used for final model prediction, but so much more data is accumulated during training. In this talk, we will discuss ways of utilizing data from the entire training trajectory in the context of image classification to improve final model performance, at a fraction of the cost of traditional ensembles.

Max Halford

Data Scientist @ Carbonfact

Online Machine Learning With RiverML

A growing number of data teams have to deal with real-time data feeds. Handling these feeds is challenging. Part of the reason comes down to habits: data processing is usually done in a batch fashion. This is very much the case for machine learning, which involves inference and learning. Both of these can be done online. But how? What design patterns does this require? What software components are necessary? How does this look in practice from day to day? We’ll try to refine these questions and answer them during this talk. In particular, we’ll focus on River, which is a Python package for online machine learning. We’ll also discuss the higher-level tools necessary to deploy an online machine learning model into production.

Claire Detilleux

Senior Data Scientist @ Spotify

Our Journey at Spotify Search Towards a More Data-Driven Product Development Culture

At Spotify, we aim to build and improve our product in a data-informed way. To do that, teams are encouraged to generate and test hypotheses by running experiments and gathering evidence for what works and what doesn’t. In the Search team, on our journey towards this goal, we have learned that, besides having the ambition, we need at least two more things:

1. An experimentation platform that allows us to run experiments at scale and generate accurate results
2. A product development culture with evidence-based hypothesis testing at its core

The talk would be about the key drivers that enabled our journey in the Search team at Spotify towards building a more data-driven culture for product development.

Martin Byröd

Engineering Manager @ Apple

Building the Next Generation Digital Maps Using a Fusion of 3D Computer Vision and Deep Learning

The past decade has seen a revolution in the ability of machines to extract high-level, semantic meaning from raw sensory data. This revolution has primarily been brought about by deep learning. In combination with exponential growth in the availability of compute and the rapidly decreasing cost of sensing, this has been an explosive change to the field of digital mapping. The 3D Vision team of Apple Maps capitalizes on these trends to push the boundaries of mapping. By using state-of-the-art technologies from machine learning and computer vision, we can redefine the modern digital map and how it is built. In the talk, we will walk you through our end-to-end philosophy for large-scale mapping in the age of deep learning. We take a look under the hood with examples and visualizations to explain how our approach has led to many breakthrough products like Look Around, Visual Localization for Geo-Referenced AR, and the new 3D Apple Maps.

Fredrik Bagge Carlson

Julia for AI and Data Science

Julia is a programming language with a set of features and a package ecosystem that makes it especially useful for AI/ML and data science. While the language is relatively young (measured in programming language years) the productivity it brings means that there is already a number of mature packages that push the state-of-the-art forward in domains such as differentiable programming, simulation, and optimization.

In this talk, we will give a high-level overview of Julia with an emphasis on features, tools, and packages that makes it suitable for a typical AI/ML and data science workflow. We will also give some examples of ongoing research done in Julia, such as differentiable programming and automatic surrogatization for increased productivity in simulation and model-based workflows.

Tommy Johansson

Perception Expert @ Annotell

It All Starts With a Consistent Labeling Guideline

The road to reliable ground truth starts before any labeling is even made. In perception software development, when training a network, the specifications of the wanted function need to be translated to the labeling guideline. The function and its desired behavior need to be described in the labeling guideline to get the expected behavior from the model in the end.

During annotation, ambiguous situations frequently pop up; these must be dealt with, preferably before starting the labeling work. A guideline shall also be understood and interpreted in the same way by hundreds of persons, usually with different cultural backgrounds and education.

How can we achieve a solid guideline that captures all function needs and has no room for ambiguities, and how can we assure that the rules/requirements are fulfilled? This presentation will show a structured way to minimize the risk of ambiguities and inconsistency in the guidelines and the tools available at Annotell to help you.

Sebastian Nabrink

Co-Founder @ Oterlu AI

Deploying Deep Learning to Prevent Online Hate

How do you deliver Natural Language Processing (NLP) services using deep learning when resources are scarce, and requirements on latency are demanding? That is the daily struggle at Oterlu AI, a Gothenburg-based tech startup setting out to tackle online hate. Listen to co-founder Sebastian Nabrink as he shares their experience with this thus far.

Francesca Carminati

Data Scientist @ Peltarion

Topic Classification for Survey Analysis with the Peltarion Platform

This presentation will go through the main steps and challenges of a real-life project, from the data collection to the end model. In this case, my team and I worked with a customer that analyses employee surveys to help the management identify problematic areas within their company. Our final solution involved fine-tuning a multilingual transformer model and individual threshold tuning for each label. The model was built, trained, and deployed on the Peltarion Platform.

Erik Stenberg

Data Scientist @ Spotify

How Spotify Analyzes Thousands of A/B Tests With Confidence

Confidence (spotify-confidence on PyPi) is an open-source Python library that has been used to analyze experiments at Spotify for more than four years. It is now also an integrated part of Spotify’s internal experimentation platform, where it’s used to automatically analyze thousands of experiments every year. In this talk, we will discuss its benefits, demonstrate how to use it, and talk about what lies ahead.

Anil Sharma

Data Science Manager @ Klarna

From Nothing to Something: Klarna’s Journey With Recommendation Systems

This talk will be about Klarna’s journey from zero recommendation models to our current state of five use cases in one year. I will discuss challenges, wins, learnings and give a sneak peek of where we are headed.

Erik Andersson

Product Manager BI and Analytics @ Västtrafik

Alert, Your Trip Will Be Crowded: Crowding Prediction at Västtrafik

When the pandemic hit the world, Västtrafik wanted a precise way of informing travelers on what trips were likely to be crowded so travelers could plan their trips safely. Since Västtrafik is working in a large geographical area and has a heterogeneous population landscape, creating warnings based only on the time of day wasn’t considered good enough. We will present a use case of machine learning in production based on data from the automatic passenger counting system. The focus will be on all those details that make ML solutions useful in real life. How do you communicate the results? How do you maintain it? What engine does the model use? We will also present results on how it was received by the travelers and how it is continuously evaluated and evolving today.

Per Sillrén

Senior Data Scientist @ Spotify

How Spotify Analyzes Thousands of A/B Tests With Confidence

Confidence (spotify-confidence on PyPi) is an open-source Python library that has been used to analyze experiments at Spotify for more than four years. It is now also an integrated part of Spotify’s internal experimentation platform, where it’s used to automatically analyze thousands of experiments every year. In this talk, we will discuss its benefits, demonstrate how to use it, and talk about what lies ahead.

Magnus Röding

Predicting Properties of Materials With ML

Porous materials and their mass transport properties (for example fluids flowing in the pores, or diffusion of particles) are essential for the functionality of products in many industries e.g. pharma, hygiene, wound care, packaging, and energy. Understanding the microstructural geometry and how it relates to mass transport properties is crucial for designing better materials. We combine e.g. image analysis, statistics, numerical simulations, and machine learning to characterize materials and to predict, understand, and optimize their properties. In this talk, we will focus on the usefulness of machine learning in this area. We discuss a number of cases where the incorporation of machine learning techniques leads to valuable new tools and insights: for semantic segmentation of imaging data, accelerated analysis of data from other physical/chemical measurement techniques, and optimizing the design of materials structures.

Kristoffer Carlsson

Software Engineer @ Julia Computing

Julia for AI and Data Science

Julia is a programming language with a set of features and a package ecosystem that makes it especially useful for AI/ML and data science. While the language is relatively young (measured in programming language years) the productivity it brings means that there is already a number of mature packages that push the state-of-the-art forward in domains such as differentiable programming, simulation, and optimization.

In this talk, we will give a high-level overview of Julia with an emphasis on features, tools, and packages that makes it suitable for a typical AI/ML and data science workflow. We will also give some examples of ongoing research done in Julia, such as differentiable programming and automatic surrogatization for increased productivity in simulation and model-based workflows.

Erik Landolsi

Privacy Aspects of High-Performance Video Analytics

Irisity develops systems for security and safety using video analytics, triggering real-time alerts based on configured rules. For many applications, data protection and personal privacy are the main customer concerns. One example is detecting patients falling to the ground in nursing homes while maintaining absolute privacy for the monitored individuals. One of the privacy aspects is around the training data. Our tech stack relies heavily on deep learning, and it is hard to get real-world training data from sensitive environments such as hospitals in large quantities. A second aspect is privacy in live systems. Our system includes features for anonymization, where all detected humans can be replaced with a low-resolution pixelized version, hiding the person’s identity. A final aspect of privacy and data protection is to support video processing at the edge, as many of our customers are reluctant to upload video data to cloud systems. We found using GPUs at the edge to be impractical in many cases, and spent significant efforts on making our model inference run highly optimized on edge CPUs. This talk describes the details on how we address these three types of challenges, with examples, results, and a few lessons learned.

Niclas Johansson

Machine Learning Engineer @ Tenfifty

Oompa Loompa and How to Save Lives Using Speech to Text

I will talk about how we at Tenfifty, together with The Swedish Maritime Administration, use AI to help the rescue leaders working at the Joint Rescue Co-ordination Centre (JRCC) save lives around the Swedish coastline. The project named Heimdall is set out to detect emergency messages sent using VHF around the Swedish coast. I will talk about how open source can outperform existing solutions and how big of an impact noisy audio and the specific domain have on existing models.

Erik Coelingh

Vice President, Product @ Zenseact

Towards Zero Faster Using Deep Learning

Description to be announced

Björn Thalén

Data Scientist @ Västtrafik

Alert, Your Trip Will Be Crowded: Crowding Prediction at Västtrafik

When the pandemic hit the world, Västtrafik wanted a precise way of informing travelers on what trips were likely to be crowded so travelers could plan their trips safely. Since Västtrafik is working in a large geographical area and has a heterogeneous population landscape, creating warnings based only on the time of day wasn’t considered good enough. We will present a use case of machine learning in production based on data from the automatic passenger counting system. The focus will be on all those details that make ML solutions useful in real life. How do you communicate the results? How do you maintain it? What engine does the model use? We will also present results on how it was received by the travelers and how it is continuously evaluated and evolving today.

Anass Sedrati

Business Technologist @ Bintess

What Is the Place of Democracy in an Intelligent Device World?

Democracy is one of the most appreciated and encouraged concepts throughout the world. One of its pillars is equality, suggesting that all humans are equal in choice, vote, and decision-making. In the future (and already now), artificial intelligence will empower countless devices to make more and more decisions based on data and machine learning. An important question emerging from this new shift will be if the concept of democracy will need to be revisited with the strong upcoming of AI?

Democracy as a concept was created to regulate human governance, given that they were the only beings reaching the intelligence level required to discuss decisions and share power. But with the advent of artificial forms of intelligence, will their existence be considered in future decision-making? Will intelligent devices have the same weight as humans or citizens with the same level of knowledge? What gives legitimacy to humankind to hold the ultimate power over other intelligent forms?

Henrik Alburg

Data Science Manager @ Monocl

Moving Monocl’s Spark Workflows to Kubernetes

Monocl collects data from various public sources within life sciences, spanning publications, clinical trials, meeting presentations, and research grants. We extract and disambiguate all the names from these sources, allowing us to figure out who has done what, making them searchable on our platform. Doing this name disambiguation can be both complicated and resource-intensive, so we use Apache Spark to distribute the load to a cluster.

From the start, we managed our very own Spark clusters using Ansible and Bash, running Spark in standalone mode. But as the data team, the weekly commitment, and the number of sources has grown, we needed a new, more scalable solution. Looking at the different alternatives, we decided to go with Spark on Kubernetes and use Argo Workflows as our workflow orchestration tool. This talk is about why those choices were made and how it all turned out.

Leonard Aukea

Head of Machine Learning Engineering & Operations @ Volvo Cars

Effective ML Systems Development