Maria Larsson

Machine Learning Engineer @ Sleep Cycle

Few-Shot Learning for Health: How Sleep Cycle Separate Multiple People’s Snoring

How do you distinguish snoring sounds from two or more people sharing a bedroom? And why is it important? With several million users choosing to share their data with Sleep Cycle, we are in a unique and privileged position to deepen our understanding of the impact snoring can have on our sleep quality and, by extension, our well-being. Our vast dataset has allowed us to train a snore embedding network that can differentiate between different people’s snoring, regardless of how many people or pets(!) are sharing a bedroom.

Snoring, often dismissed as a nuisance only to the person sharing the room, can be a sign of obstructive sleep apnea and is linked to a wide range of other negative health-related issues. Providing the user with a comprehensive sleep analysis and letting them know where the snoring is coming from is the first step for them to implement the proper measures for their sleep health and overall well-being.

Zenodia Charpy

Senior Data Scientist @ NVIDIA

P-Tuning: A Parameter Efficient Tuning to Boost LLM Performance

As more LLMs become available, industries need techniques for solving real-world natural language tasks. It has been shown that model prompting methods can elicit good zero– and few-shot performance from LLMs and help yield quality results on various downstream natural language processing (NLP) tasks. However, there is a limit to it. In this talk, we will demonstrate how to adapt p-tuning, a prompt-learning method, to low-resource language settings. We use an improved version of p-tuning implemented in NVIDIA NeMo that enables the continuous multitask learning of virtual prompts. In particular, we focus on adapting our English p-tuning workflow to Swedish.

Salman Toor

CTO @ Scaleout Systems

Federated Machine Learning: A Scalable, Privacy-Preserving Approach Ready for Production Environments

Federated machine learning has created new possibilities for privacy-preserving data analysis. This is a thriving area of research that primarily focuses on the algorithmic details and communication overheads necessary to train accurate models. Despite significant progress in the field, production-grade federated machine learning frameworks that address essential properties such as security, data privacy, scalability, fault tolerance, and performance in geographically distributed settings have yet to be available to ML engineers.

This talk will highlight the core concepts and features necessary for developing federated learning platforms for production environments at Enterprises. Furthermore, it will briefly cover two active use cases demonstrating the potential of regulated datasets in geographically distributed settings.

Laszlo Sragner

Founder @ Hypergolic

Clean Architecture: How to Structure Your ML Projects to Reduce Technical Debt

Software engineering principles are frequently mentioned as a solution to data science’s productivity problem. Unfortunately, rarely in a comprehensive format to be actionable or adopted for data-intensive use. In this talk, I will present the Clear Architecture framework that enables practitioners to structure their projects and manage changes throughout their lifecycles. The audience will also learn about a minimum set of programming concepts to make this a reality. The key takeaway is that, as a data scientist, you can take care of your codebase with only a few techniques and a little effort.

Christian Berg

Founder @ Paliscope

Finding the Bad Guys with AI: Connecting the Dots In Data

Paliscope is a Swedish company developing software for analyzing large amounts of data like tables, images, text, audio, and video with analysis methods like object detection, scene classification, face recognition, natural language processing, and speech-to-text. Since everything is connected somehow, finding the links, whether they stem from chat logs, speech, or social media posts, is essential. AI is critical in connecting the dots as everyone otherwise would be overloaded with information. Paliscope products are primarily used by law enforcement and other government agencies in several countries. During the presentation, we will demonstrate a couple of use cases for how the software Paliscope YOSE is being used and focus on a couple of technologies used within the tool.

Alma Lund

Machine Learning Engineer @ Mabel AI, Syntronic

Using AI to Preserve Privacy in Healthcare

Language barriers in health care are a major obstacle to national and global sustainable development goals for public health, health equality, and integration. Interpreters in healthcare is a non-regulated industry, with less than 1 % of the interpreters being certified in medical interpretation in Sweden. Often the interpreter knows the patient, making it difficult for the patient to have a private conversation with the doctor. In this talk, we will look at how modern NLP and speech technology can be used to create a secure and private interpretation system. We will go through our challenges in making such a system with the first version designed to help the Ukrainian refugees seeking healthcare across the EU. We will particularly look into dealing with noisy environments and biases in open-source datasets.

Francisca Hoyer

Strategic Program Manager, NLU @ AI Sweden

An Interdisciplinary Expert Pool for Natural Language Understanding

The artifacts used as training data in Machine Learning have their own historical, social, and cultural contexts – the majority of them were not produced to be used as training data for machine learning to make future predictions. Instead, they embody, represent, and reinforce a broad range of ideas and values, power relations, stereotypes and prejudices, and systematic forms of discrimination situated in their own places in time and space. Large Language Models “learn” languages by reading large amounts of these human-produced texts. Thereby they also “learn” the ideas and values, power relations, stereotypes and prejudices, systematic forms of discrimination, etc., inscribed in the training data. Therefore, questions of data, representativeness and fairness, equality, and ethics must be central to the development of language technologies. In this presentation, AI Sweden presents a new form of cross-disciplinary collaboration that aims to tackle these issues in concrete use case scenarios: The interdisciplinary expert pool for Natural Language Understanding, a platform for knowledge exchange between humanities scholars, social scientists, civil society organizations, and AI development teams.

Jussi Karlgren

Docent in Language Technology @ Helsinki University

Should We Teach Machine Learning Systems?

When we find that AI-powered systems behave unexpectedly or inappropriately, we often discuss how we can retrain and find or generate better or unbiased data. This is because we have a general hands-off perspective on learning: we expect a system to build a better model if the learning process is left to make sense of input data with minimal supervision and guidance.

This talk will discuss the place of feature engineering and instruction in machine learning to mitigate some of the observed challenges of providing users with high-quality and reliable output and contrast it with today’s approach.

Magnus Sahlgren

Head of Research, NLU @ AI Sweden

GPT-SW3: The First Large Generative Language Model for the Nordic Languages

This talk gives an overview of the process of building the first large generative language model for Swedish. We cover the motivation for building the model, as well as challenges and opportunities with data and compute. We also give examples of applications of the model and discuss future directions for building and deploying large language models for smaller languages.

Marina Delatorre

Data Scientist @ Mabel AI

Using AI to Preserve Privacy in Healthcare

Language barriers in health care are a major obstacle to national and global sustainable development goals for public health, health equality, and integration. Interpreters in healthcare is a non-regulated industry, with less than 1 % of the interpreters being certified in medical interpretation in Sweden. Often the interpreter knows the patient, making it difficult for the patient to have a private conversation with the doctor. In this talk, we will look at how modern NLP and speech technology can be used to create a secure and private interpretation system. We will go through our challenges in making such a system with the first version designed to help the Ukrainian refugees seeking healthcare across the EU. We will particularly look into dealing with noisy environments and biases in open-source datasets.

Zeeshan Dar

Senior Machine Learning Engineer @ Prisjakt

Navigating the Challenges of Offer Categorization in E-commerce: Strategies for Accurate and Scalable Classification

The accurate categorization of offers is a prominent challenge in the ecommerce industry. Effective offer categorization is critical to ensuring that products are displayed to the right customers and can provide valuable insights into customer behavior and preferences. However, categorizing offers at scale can be especially challenging when dealing with large product catalogs. To address this, it is important to use a robust system that can accurately categorize offers based on factors such as product attributes, price, and to leverage automation tools such as machine learning algorithms where possible.

Offer categorization can also help with search relevance, sorting offers to products, and facilitating easier indexing by search engines such as Google. Mapping offers to Google’s taxonomy, which comprises of nearly 6000 hierarchical classes, can quickly become challenging when working at scale. To ensure consistent categorization across different systems and platforms, it is important to establish clear taxonomy standards and to continually review and adjust the categorization system to ensure that it remains accurate and effective.

In this talk, we will discuss the challenges encountered in categorizing offers, the strategies we employed to address scaling and categorizing concerns, as well as the issues and limitations we overcame to achieve success.

Mattias Ulmestrand

Technical Developer @ Smart Eye

Vision-Based Activity Detection Inside the Vehicle

As distracted drivers cause many road fatalities, activity detection is crucial to automotive in-cabin sensing. In today’s cars, cameras are used to monitor the driver’s gaze. But while gaze aversion is a critical factor of driver distraction, studies show that the risk increases based not only on what you look at but how you engage.

Robust activity detection in a car requires the fusion of a multitude of signals that each have their unique strength. For example, body movement over time can show that a person is engaged in a secondary, non-driving-related task. To further understand the scene, this will have to be combined with an appearance-based algorithm, such as a deep convolutional neural network that detects interaction with objects. On top of this whole cabin view, the well-established driver monitoring signals can further strengthen insight into occupant activity, such as detecting speaking by analyzing lip movement or gaze direction, which can indicate talking or texting on the phone.

This talk will give insight into the steps in developing such an activity detection system – from data collection and annotation to integration to a System on Chip. Examples will be presented in addition to some useful tips and tricks when developing various detection algorithms.

Matteo Molteni

Data Engineer @ Kognic

Leveraging Open-Source Tools for Building a Quality Data Warehouse

In recent years the modern data stack has seen a great increase in the number of tools available on the market, making the landscape overwhelming and difficult to navigate. In this presentation, we will give an overview of how Kognic built a data warehouse with quality as its core value. We will present our architecture and explain how we leverage some popular open-source tools in our setup, highlighting their main features and shortcomings. Particular emphasis will be put on how data is transformed for the data mart using a test-driven approach that aims to stay true to software-development best practices.

Kokchun Giang

Educator @ IT-Högskolan

Building a Data Science Program in Higher Vocational Education

As the industry is becoming more data-driven, hiring roles with more data-related skill sets, such as data scientists, data engineers, and data analysts, is necessary. To meet this need in the industry, we have built a two-year data science program in higher vocational education together with industry experts. The program contains courses in programming, data processing, mathematics, statistics, machine learning, deep learning, data engineering, and more. Besides these theoretical courses, there are two internship courses where students apply their skills to real data science problems in different companies.

The program is also built using a modern approach to education that follows the trend in society. For example, we have created a Discord community for each cohort, alum, and industry partner. In this way, the students can share experiences and knowledge and help each other, becoming each other’s learning resources. The students also use GitHub throughout their education, building portfolios to share with future employers.

Mats Kvarnström

Head of AI @ Paliscope

Finding the Bad Guys with AI: Connecting the Dots In Data

Paliscope is a Swedish company developing software for analyzing large amounts of data like tables, images, text, audio, and video with analysis methods like object detection, scene classification, face recognition, natural language processing, and speech-to-text. Since everything is connected somehow, finding the links, whether they stem from chat logs, speech, or social media posts, is essential. AI is critical in connecting the dots as everyone otherwise would be overloaded with information. Paliscope products are primarily used by law enforcement and other government agencies in several countries. During the presentation, we will demonstrate a couple of use cases for how the software Paliscope YOSE is being used and focus on a couple of technologies used within the tool.

Kajsa Andreasson

Software Engineer @ Recorded Future

Cross-modal Transfer Between Vision and Language for Protest Detection

Multimodal data (data with two or more modalities like text, images, or audio) has gained more and more attention in the last year. One example is the image-generating model DALL·E 2, released by OpenAI, which uses the modalities of images and text. Even though multimodal models have proven to have great potential, most of today’s systems for socio-political event detection are text-based.

In this presentation, we discuss a proposed approach of using the increasing amount of multimodal data to decrease the need for annotation – as presented in our paper “Cross-modal Transfer Between Vision and Language for Protest Detection.” We propose a method that utilizes existing annotated unimodal data to perform zero-shot event detection in another data modality. Specifically, we focus on protest detection in text and images and show that a pretrained vision-and-language alignment model (CLIP) can be leveraged to this end. In particular, our results suggest that annotated protest text data can act supplementarily to detect protests in images, but significant transfer is also demonstrated in the opposite direction.

Anders Hildeman

Senior Data Scientist @ AstraZeneca

Unlocking Hearts by Locking Models: Solving Multiple Echocardiogram Problems Using a Holistic Self-Supervised Approach

Supervised training of deep learning models for medical imaging applications requires a significant amount of labeled data, which can be challenging to acquire as it necessitates the expertise of highly skilled medical professionals to annotate a large number of image sequences. To overcome this limitation, we propose the Adaptive Locked Agnostic Network (ALAN) concept for label-free, self-supervised learning of visual features. The advantage of this methodology is that the self-supervised training can be performed once on a large unlabeled dataset, and the features discovered can then be used on a plethora of subsequent downstream tasks. Since these self-engineered features have rich informational content, downstream tasks can be solved with more straightforward and less data-hungry methods than in fully-supervised approaches. In this paper, we perform two typical downstream tasks for echocardiogram data: segmentation of anatomical regions and classification of the echocardiogram views. We evaluate the proposed self-supervised approach on three publicly available echocardiogram datasets: EchoNet-Dynamic, CAMUS, and TMED-2.

Abgeiba Isunza Navarro

Machine Learning Engineer @ Modulai

Synthetic Data for Computer Vision Problems, When Real Annotation Is Too Expensive

Deep learning has been successfully used for various computer vision problems. However, models often need large amounts of data to solve a specific task effectively. Unfortunately, there is often a bottleneck to accessing labeled data. With the advancement in graphic tools, simulator engines, and generative models, synthetic data generation has become a suitable alternative to approach the lack of data in specific domains.

We will talk about the current development in synthetic data generation and how we have used it to build good models for computer vision tasks based primarily on synthetic data. We will go through the different use cases, explain the tools used and give an insight into how we approach a computer vision problem when we lack a large amount of labeled data.

Ria Raj

Software Engineer @ Recorded Future

Cross-modal Transfer Between Vision and Language for Protest Detection

Multimodal data (data with two or more modalities like text, images, or audio) has gained more and more attention in the last year. One example is the image-generating model DALL·E 2, released by OpenAI, which uses the modalities of images and text. Even though multimodal models have proven to have great potential, most of today’s systems for socio-political event detection are text-based.

In this presentation, we discuss a proposed approach of using the increasing amount of multimodal data to decrease the need for annotation – as presented in our paper “Cross-modal Transfer Between Vision and Language for Protest Detection.” We propose a method that utilizes existing annotated unimodal data to perform zero-shot event detection in another data modality. Specifically, we focus on protest detection in text and images and show that a pretrained vision-and-language alignment model (CLIP) can be leveraged to this end. In particular, our results suggest that annotated protest text data can act supplementarily to detect protests in images, but significant transfer is also demonstrated in the opposite direction.

Christoffer Petersson

AI Researcher @ Zenseact

Deep Learning for Self-Driving Cars

The mission of Zenseact is to develop a world-leading autonomous driving software platform for consumer vehicles, with the primary goal of dramatically reducing the number of traffic fatalities and injuries around the world. I will discuss how we use deep learning to reach this goal, some of the key challenges we see, and how we plan to expand the use of learning-based algorithms. I will also review some of our recent and ongoing deep learning-related research activities.

Jens Agerberg

ML Researcher, Data Scientist @ KTH, Ericsson

Teaching Neural Networks a Sense of Geometry

By taking neural networks back to the school bench and teaching them some elements of geometry and topology, we can build algorithms that can reason about the shape of data. Surprisingly these methods can be useful not only for computer vision but in a wide range of applications, such as evaluating and improving the learning of embeddings or the distribution of samples originating from generative models. This is the promise of the emerging field of Topological Data Analysis (TDA) which we will introduce and review recent works at its intersection with machine learning.

Claes Strannegård

Associate Professor @ Department of Applied IT, University of Gothenburg

Modelling Ecosystems with Deep Reinforcement Learning

I will present the ecosystem simulator Ecotwin (www.ecotwin.se), which uses three-dimensional terrain models based on real geographical data together with individual animal models controlled by deep reinforcement learning. Moreover, I will illustrate how this simulator can be used for studying pristine ecosystems as well as ecosystems affected by human economic activities such as exploitation of natural resources, land use change, pollution, and climate change. The main benefit of the simulator is that it can be used as an analytical tool at the planning stage for understanding local ecological consequences of a range of economic activities, thus ensuring relatively eco-friendly decision-making.

Kajsa Norin

NLP Consultant @ Kajsa Norin AB

Training and Using a Large Language Model to Write Theatrical Performances

Will all writers have their own AI in the future? This project explores the possibility of training a large language model on a writer’s previous material in order to create an AI that acts as a creative assistant. In recent years, there has been a growing interest in using artificial intelligence to generate creative content. One such application is the use of large language models (LLMs) to write theatrical performances.

In this talk, we will review the technical aspects of training and using an LLM to generate scripts for theatrical performances. We will discuss the challenges of working with natural language data, the techniques used to preprocess and clean the data, and the methods used to train and fine-tune the LLM as well as deploying it. A performance written by the model and writer in collaboration will premiere in Stockholm at Dansens Hus in November 2023.