Zenodia Charpy

Senior Deep Learning Solution Architect | NVIDIA

Better, Faster utilizing your GPUs for deep learning workload

The utilization of GPUs has evolved from gaming purposes, to deep learning, and gear towards, but not limited to, computer vision type of workloads :autonomous vehicles, medical images in radiomics and video surveillance to name a few.

However, having powerful GPUs at your disposal is one thing, knowing how to use them to their full potential is something we are going to explore together.

Working at Nvidia, we are dedicated to create an eco-system with full cycle of hardware-and-software solutions, that enables optimization in all fronts: from providing plug-and-play-able docker repo maintained by us, to data augmentations within GPU on-the-fly, to parallel model training using multiple gpus with mixed_precision for target deployment.

We have open-sourced all these toolkits for data scientists to quickly kick-start and get to work, instead of spending precious time to fix environmental installation errors and bring super computing power at your fingertips.

Today we are going to take a look at these essential toolkits which enable deep learning practitioners, data scientists alike to improve their model training in parallel, iterate faster and develop better market-ready products.

Thore Husfeldt

Professor of Computer Science | Lund University

An Introduction to Algorithmic Fairness

I will give a brief introduction to algorithmic fairness. I will introduce and compare several different definitions of fairness that can be used to evaluate outcomes of algorithmic decision procedures and highlight their relationship. Of particular interest are the tensions between notions of individual and group fairness, and between various notions of group fairness. The motivation and main take-aways from the presentation are aimed at a general academic audience, but I will include mathematically precise terms.

Robert Feldt

Professor of Software Engineering | Chalmers

Testing of Machine Learning Systems – the importance of “lagom” surprising inputs

Testing of machine learning (ML) components, such as deep neural nets, is not only about correctness and accuracy; there is a large number of quality properties that has to be ensured. While research on how to perform these different forms of testing is still immature it is growing at a tremendous pace. In this talk, I’ll give an overview of recent results on testing ML models and discuss how it differs from testing normal software. I will exemplify with recent work on how to find adequately (“lagom”) surprising test inputs, that are not random or look like noise, but rather that are realistic. While the current focus of research is mainly on neural nets I’ll also discuss if and how this might be generalised to other types of machine learning models.

Rocío Mercado

Postdoctoral Fellow | AstraZeneca

Discovering New Molecules Using Graph Neural Networks

There is growing interest in graph neural networks (GNNs) for graph representation learning. This is because graphs are natural data structures for describing an assortment of relational information, including molecular structures. One of the topics we are interested in within the Molecular AI (MAI) group at AstraZeneca is using GNNs for the design and discovery of new drug molecules.

The process of designing novel, drug-like molecules can be viewed as one of generating graphs which optimize all the features of the desirable molecules. We are interested in graph-based methods because they have the potential to capture a lot of molecular information with greater flexibility. Our generative models can quickly learn the underlying distribution of properties in training set molecules without any explicit writing of chemical rules.

In this talk, I will demonstrate how we can use deep learning methods such as GNNs to carry out pharmaceutical drug  discovery more efficiently.

Oscar Söderlund

Chief Software Architect | Einride

Scalable forecasting in Google Cloud

A practical and data engineering-oriented talk on how to stand up a scalable and fully automated pipeline for time series forecasting using Facebook’s Prophet library and the standard big data and machine learning tools in Google Cloud.

This will be done in the context of Einride’s data platform, which is all about creating actionable insights that drive customers toward sustainable transport. As a real-world case-study, I will show how we break down and understand transport demand in multiple dimensions – from a customer’s total demand down to thousands of individual sites and shipping lanes.

I will walk us through the full pipeline; extracting a production database dump, multiple tiers of data cleaning and transformation, how to use PySpark and Dataproc to parallelize model training and forecast generation, and how to orchestrate it all with Apache Airflow.

The key takeaway (and what’s really exciting) is how easy to use and available big data and machine learning tools have become – to tech giants and fledgling startups alike!

Pablo Estevez Castillo

Principal Data Scientist | Booking.com

What we have learned from 150 successful ML-enabled products at Booking.com

Booking.com is the world’s largest online travel agent where millions of guests find their accommodation and millions of accommodation providers list their properties including hotels, apartments, bed and breakfasts, guest houses, and more. During the last years we have applied Machine Learning to improve the experience of our customers and our business. While most of the Machine Learning literature focuses on the algorithmic or mathematical aspects of the field, not much has been published about how Machine Learning can deliver meaningful impact in an industrial environment where commercial gains are paramount. We conducted an analysis on about 150 successful customer facing applications of Machine Learning, developed by dozens of teams in Booking.com, exposed to hundreds of millions of users worldwide and validated through rigorous Randomized Controlled Trials. Following the phases of a Machine Learning project we describe our approach, the many challenges we found, and the lessons we learned while scaling up such a complex technology across our organization. Our main conclusion is that an iterative, hypothesis driven process, integrated with other disciplines was fundamental to build 150 successful products enabled by Machine Learning.

Olof Mogren

Machine Learning researcher | RISE

Social bias and fairness in natural language processing

Learned continuous representations for language units was the first trembling steps of making neural networks useful for natural language processing (NLP), and promised a future with semantically rich representations for downstream solutions. NLP has now seen some of the progress that previously happened in image processing: the availability of increased computing power and the development of algorithms have allowed people to train larger models that perform better than ever. Such models also make it possible to use transfer learning for language tasks, thus leveraging large widely available datasets.

In 2016, Bolukbasi, et.al., presented their paper “Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings”, shedding lights on some of the gender bias that was available in trained word embeddings at the time. Datasets obviously encode the social bias that surrounds us, and models trained on that data may expose the bias in their decisions. It is important to be aware of what information a learned system is basing its predictions on. Some solutions have been proposed to limit the expression of societal bias in NLP systems. These include techniques such as data augmentation and representation calibration. Similar approaches may also be relevant for privacy and disentangled representations. In this talk, we’ll discuss some of these issues, and go through some of the solutions that have been proposed recently.

Rebecka Jacobsson

Machine Learning Lead | iZettle

Building a feature library for machine learning at iZettle

When we started building features for our first machine learning models at iZettle, we quickly realized we didn’t want to reinvent the wheel for every model we built. Instead, we created a shared feature building framework, which over time has evolved into a core part of our machine learning infrastructure. In this talk, I will explain why you might want to create a feature library, suggest some important trade-offs to consider and share how we first built the simplest solution possible and then iterated from there.

Daniel Gillblad

Head of AI Research | RISE

AI – Perspectives and Strategies

While AI and its applications continue to develop rapidly, many challenges remain to realise the full potential. We will reflect over the developments within AI and Machine Learning during the last few years, the current state-of-the-art, what remains to be done and what some of the urgent issues are at the moment. With this as a starting point, we will discuss the current state of AI and AI applications in Sweden and how we can think about national strategies for accelerating the use of AI nationally, in both industry and public sector. Finally, we will provide some examples of ongoing work in important areas.

Christian Forssén

Professor in Theoretical Physics | Chalmers

Learning from data in fundamental physics research

Bayesian statistics allows to quantify the strength of inductive inference from facts (such as experimental data) to propositions such as scientific hypotheses and models. It can be used for example in searches for physics beyond the standard model, and in multi-message astronomy for the analysis of gravitational wave signals. In our theoretical-physics research we develop novel machine-learning methods to make the Bayesian approach tractable for models that involve heavy computing.

A general message of this presentation will be that the Bayesian philosophy can be embraced by machine-learning practitioners to introduce uncertainties and to avoid overfitting.

Daniel Langkilde

Co-founder | Annotell

The impact of data quality on model performance

A lot of focus in the machine learning community is on the choice of algorithm. This talk will instead focus on the dataset. While most people by now have realized that dataset quality is critical for the success of machine learning projects, most teams still struggle to quantify data quality. This talk will provide hands-on examples of how to quantify dataset quality, and also reason about the impact various types of quality issues can be expected to have on model performance.

Martin Tegner

Machine learning researcher | IKEA Digital

Data privacy and fairness in recommender systems

In early 2020, IKEA made a promise to give back the control to customers over their data. In this talk, we look at two aspects of the promise in the context of data-driven product recommendations. First, we analyse performance of such systems under minimal data requirements. Second, we propose a Bayesian approach to recommendations that uses in-session active learning to give personalised content without collection of private data.

Adam Andersson

Radar Systems Engineer | Saab Surveillance

Removing computational bottlenecks with deep learning

Deep learning has transformed the fields of computer vision and natural language processing. There is a more recent trend in using neural networks to solve differential equations. It is less well known and has so far mostly gained academic interest. Computer aided engineering is, under the hood, based on solving partial differential equations (PDE) approximately with classical methods such as the finite element method.

A crash simulation takes hours to run on a computer cluster and the PDE has 4 dimensions (three space dimensions and one time dimension). Last year, various nonlinear PDE with 10000 space variables were solved approximately with deep learning on much smaller computers. The problems are not perfectly comparable but this result still indicates the ability of deep learning to scale beyond classical methods. Both automatic control and sensor fusion problems can be formulated and solved in terms of PDE, but due to computational complexity of classical methods these are not used in practical algorithms. In this talk I discuss how such equations for stochastic control and potentially for sensor fusion can be used for deriving practical algorithms, with deep learning being the enabler.

Carl Lindberg

Co-Founder / Director of Strategic Data Applications | Sigmastocks / Zenseact

Common Pitfalls when Working with Data

With the present AI and Data Science hype, the marvelous theory of Statistics has found yet two more fields to be applied to. However, the Statistical Machine works differently than most other machines you have heard of, and there are many pitfalls into which Data Scientists innocent of Applied Statistics too easily fall. This talk will deal with some of the most common such pitfalls, illustrated by relevant real life examples.