Allison Cohen

Senior Applied AI Projects Manager @ AI for Humanity, Mila – Quebec AI Institute

Why Take a Unidimensional Approach to Tools That Exist in High Dimensional Space?

There is a problematic and yet ubiquitous phenomena in the field of AI research and product development: we’re too late to ask ourselves whether the algorithms we’ve been developing are built for purpose. Why? Because we’ve been misinterpreting “fit for purpose” to mean technically feasible. This definition obscures considerations of equal importance including the cultural, social, political and legal landscape that permeate the tool’s design and determine the tool’s utility.

When building AI products, researchers are only confronted with questions of multidisciplinary reflection as they’re about to submit a paper or launch their tool. However, at this point in the project, a host of relevant decisions have already been made, whether consciously or not, that influence the algorithms’ utility. These decisions begin before the algorithm has ever been trained or the data has ever been collected.

In this talk, I will discuss important points of inflection that are too often missed in the product development lifecycle. These inflection points present opportunities for AI researchers and product developers to ensure that the technology they’re building is fit for purpose.

Charlotta Kronblad

Dr in digital transformation @ Gothenburg University

Daniel Gillblad

Chief of AI @ Recorded Future

Jair Ribeiro

Analytics and Insights Leader @ Volvo Group

Julia Lemonte

AI Governance Manager @ AstraZeneca

Panel Discussion: Navigating AI’s Labyrinth: Ethical and Legal Threads

Sara Wrige

Parish priest in the Church of Sweden

Panel Discussion: Navigating AI’s Labyrinth: Ethical and Legal Threads

Mina Alibeigi

AI research lead and scientist @ Zenseact

Zenseact Open Dataset for Advancing Autonomous Driving

In this talk, we will explore the critical role of comprehensive and varied driving data in the development of reliable and effective autonomous driving technology. A key focus will be on the Zenseact Open Dataset (ZOD), a large, diverse, multimodal dataset recently released to propel autonomous driving advancements. We will delve into the unique features of ZOD, outlining how it was created and the measures taken for its public release. This release aims to spur innovation in the field, pushing forward robust and safe autonomous driving while enhancing overall road safety. Additionally, the talk will briefly highlight our recent and ongoing research projects, showcasing how ZOD is being utilized in various potential applications.

Aron Lagerberg

Principle AI Scientist @ Recorded Future

Retrieval Augmented Generation for Threat Intelligence

The previous years’ advances in generative AI and Large Language Models (LLMs) have completely revolutionized the way the world thinks about AI and what problems it can solve. LLMs have been positioned as an interface to data and knowledge in various shapes or form, making knowledge queryable by natural language. In practice however, this comes with many critical challenges: how can you leverage real-time data and not just what the LLM was trained on? How can you make the answers reliable? How can you make such a system fast enough? In this talk, we will dive into how we at Recorded Future have built a retrieval augmented generation (RAG) system using LLMs to be used as an interface to our threat intelligence graph. The end goal of this system is for a user to get reliable, fact-based answers to questions posed in natural language concerning threat intelligence. In particular we will discuss how we approached the above fundamental questions.

Filip Granqvist

Machine Learning Research Engineer @ Apple

Advancing Private Federated Learning: Insights and Innovations from Research at Apple

Private Federated Learning (PFL) is an approach to collaboratively train a machine learning model between edge devices with coordination by a central server, whilst preserving the privacy of the data on each edge device. PFL is an emerging field, with exponential growth in the number of papers published over the past few years. Several big tech companies invest heavily in the practical applications of PFL. 

In this talk, we will introduce Federated Learning as a subject and describe techniques to preserve the privacy of participating edge devices. We will discuss the unique challenges encountered in PFL, and advocate for further research in areas of great impact for PFL. We will briefly outline the methods Apple has chosen to implement a privacy-preserving Federated Learning system and present key results from deployed real-world applications and our published research.

Finally, we will present how to get started doing research in PFL using simulations with open source tools.

Adam Tornhill

CTO @ CodeScene

Building the Future: Exploring the Fact-based Realities of AI-Assisted Coding

Large Language Models have enabled machines to write code. The resulting movement, AI-assisted coding, promises to improve developer productivity. However, AI-assisted coding is still in its infancy. This implies that we should embrace it with caution, guardrails, and set realistic expectations beyond any marketing hype.

In this talk, Adam presents both the short- and long-term implications of using AI assistants to write code. We do so based on extensive CodeScene research analyzing over 100k AI refactorings in real-world codebases. Based on this data, we debunk the productivity claims of today’s AI assistants; it’s easy to mistake code-writing speed for productivity.

We then step out on a new path, showing how the same line of research introduces a revolutionary technology for supporting auto-generated code improvements. Using real-world demos, you will see the power of AI-assisted coding without the risks as we automatically improve existing code. In conclusion, we explore how these novel tools not only address industry challenges such as technical debt but also underscore the growing significance of comprehending code over mere writing in the age of AI. Join in!

Olena Nahorna

Analytical Linguist @ Grammarly

Evaluating LLM-based Solutions: Overcoming Challenges and Developing Effective Metrics

The introduction of LLMs has made solving complex tasks more widely available for companies and significantly expedited the development process. Ideas that previously would have taken months to implement can now be developed much faster. However, as tasks evolve in nature and pace, new challenges emerge. Specifically, the more diverse and human-like output a model can produce, the more difficult it is to assess its quality. In this talk, we will discuss the changing nature of tasks, the challenges this shift poses, and how to build an effective evaluation framework for your LLM-based project.

Edgar Bahilo Rodríguez

Lead AI Engineer @ B2-Impact

Unveiling Precision: A Novel Machine Learning Framework for Accurate Probability Estimates in Financial Industries

Nowadays, probability prediction models play an essential role in many industries. To cite only a few examples, these applications can range from models in the engineering sector, meteorology (weather forecasting) to modeling frameworks in the financial sector such as the prediction of the number of claims in insurance, credit scoring or propensity to pay forecasts.

Probability calibration is typically assessed graphically via reliability diagrams that plot an estimated version of the conditional event probability (CEP) against the forecast value, with deviations from the diagonal suggesting lack of calibration. Classical approaches to estimating CEP rely on binning and counting and have been hampered by ad hoc implementation decisions, instability under unavoidable choices regarding binning, and inefficiency.

This session shows a practical implementation of alternative calibration metrics applied to the debt management industry, where having a good calibration of probabilities can be critical and can significantly lead to increased cash flows (e.g. in the hundreds of thousands or millions of euros). An advanced machine learning framework is designed to address the proper scoring rules for the Hyperparameter Optimization (HPO) and the use of the ECCE-MAD and ECCE-R scores to have the best model selection in terms of calibration.

Abel Sancarlos González

Data Scientist @ B2-Impact

Unveiling Precision: A Novel Machine Learning Framework for Accurate Probability Estimates in Financial Industries

Nowadays, probability prediction models play an essential role in many industries. To cite only a few examples, these applications can range from models in the engineering sector, meteorology (weather forecasting) to modeling frameworks in the financial sector such as the prediction of the number of claims in insurance, credit scoring or propensity to pay forecasts.

Probability calibration is typically assessed graphically via reliability diagrams that plot an estimated version of the conditional event probability (CEP) against the forecast value, with deviations from the diagonal suggesting lack of calibration. Classical approaches to estimating CEP rely on binning and counting and have been hampered by ad hoc implementation decisions, instability under unavoidable choices regarding binning, and inefficiency.

This session shows a practical implementation of alternative calibration metrics applied to the debt management industry, where having a good calibration of probabilities can be critical and can significantly lead to increased cash flows (e.g. in the hundreds of thousands or millions of euros). An advanced machine learning framework is designed to address the proper scoring rules for the Hyperparameter Optimization (HPO) and the use of the ECCE-MAD and ECCE-R scores to have the best model selection in terms of calibration.

Ian Ozsvald

Chief Data Scientist @ Mor Consulting

Pandas 2, Dask or Polars? Quickly Tackling Larger Data on a Single Machine

Pandas 2 brings new Arrow data types, faster calculations and better scalability and even GPU acceleration in Pandas with CuDF is possible. Dask scales Pandas across cores and recently released a new “expressions” optimization. Polars is a new competitor to Pandas designed around Arrow with native multicore support. Which should you choose for modern research workflows? We’ll solve a “just about fits in ram” data task using the 3 solutions, talking about the pros and cons so you can make the best choice for your research workflow. You’ll leave with a clear idea of whether Pandas 2, Dask or Polars is the tool to invest in and how Polars fits into the existing numpy-focused ecosystem.

Do you still need 5x working RAM for Pandas operations (probably not!)? Can Pandas string operations actually be fast (sure)? Since Polars uses Arrow data structures, can we easily use tools like Scikit-learn and matplotlib (yes-maybe)? What limits do we still face? Could you switch to experimenting with Polars?

Remi Tachet des Combes

Senior Applied Scientist @ Wayve

Lingo & GAIA: Transforming Autonomous Driving with Large Language Models and Generative AI

In this talk, I will present LINGO and GAIA, two recent models developed at Wayve. LINGO is a pioneering open-loop driving commentator that utilises natural language processing to interpret and articulate driving scenes. Its unique “show and tell” feature employs referential segmentation to visually highlight areas of interest within a scene, enhancing the model’s interaction with its environment. This capability significantly advances autonomous vehicles (AV) by improving accuracy in describing surroundings, addressing model hallucinations, and bolstering safety communication. By integrating vision, language, and action, LINGO represents a critical step towards Vision-Language-Action Models (VLAMs), aiming for human-like communication and trustworthiness in AV technology.

GAIA is an advanced generative world model designed to simulate realistic driving scenarios. It leverages video, text, and action inputs to build representations of the environment and its future dynamics, enhancing AV decision-making and safety. Comprising an image encoder and a 6.5 billion parameter autoregressive transformer trained on extensive driving data, GAIA showcases scalability and superior video generation quality. Its ability to adjust scene features like weather and time, alongside predicting diverse futures and interactions with other agents, positions GAIA as a valuable tool for AV development.

Alisa Devlic

Senior Research Scientist @ Sony AI

From Hypothesis to Reality: Designing a Superhuman Racing AI Agent Using a Deep Reinforcement Learning

This work started out as a grand challenge to create an AI agent that could beat the world’s best Gran Turismo (GT) drivers. In order to develop an agent capable of competing against the world’s best drivers, GT Sophy was trained to master the following driving skills: race car control, racing tactics, and racing etiquette. This talk will tell you a story about the technical evolution of GT Sophy from a research outcome to an in-game feature that could be introduced as part of the Gran Turismo 7 (GT7) PlayStation racing game. Sony AI, in collaboration with Polyphony Digital and Sony Interactive Entertainment, designed novel deep reinforcement learning approaches with unique training and evaluation methods on a modern cloud computing platform to accommodate this project. In less than two years since GT Sophy appeared on the cover of Nature, the breakthrough AI agent has now become a permanent in-game feature of GT7.

Ekaterina Sirazitdinova

Senior Data Scientist @ Nvidia

Multimodal Generative AI Demystified

Multimodal generative AI has recently seen significant advancements, enabling the creation of realistic images, videos, and audio from textual or other inputs. However, due to the complexity of these models, understanding how they function and how to apply them in practical settings can be challenging. During this talk, Ekaterina will shed light on the inner workings of multimodal generative AI models by discussing key concepts and techniques used in their development. She will also explore various applications and use cases of this technology. The talk is intended for anyone interested in the current state of AI and its potential to produce realistic and immersive multimedia experiences.

Samuel Genheden

Associate Director @ AstraZeneca

Do Chemformers Dream of Organic Matter? Evaluating Transformer Models for Synthesis Prediction in the Pharmaceutical Domain

Language models like transformers have found a natural place in drug discovery, solving tasks such as property prediction, molecular optimization, and reactivity predictions. Transformer models trained on public data for synthesis prediction tasks, such as product and retrosynthesis prediction, have proven effective and sometimes outperform other approaches, including template-based retrosynthesis. In this contribution, we will outline our efforts of training and introducing transformer models for synthesis predictions into our production platform for synthesis planning that is used daily by chemists. We will discuss particular challenges faced when training on a large corpus of reaction data, comparison with existing models currently used by chemists for product prediction and retrosynthesis. We will show that transformer models trained on a diverse set of reactions can surpass existing models with impressive performance. Finally, we will outline outstanding issues preventing the full adoption of transformer models for synthesis prediction.

Johan Östman

Research Scientist @ AI Sweden

Information Leakage of Neural Networks

As machine learning is becoming a cornerstone of society, it increasingly encounters the challenge of handling sensitive data. This issue is magnified when trained machine learning models are shared with external entities, e.g., open-source or via API, which raises the critical question: Can sensitive information be extracted from the shared models?
In this talk, we will navigate the fascinating domain of information extraction attacks targeting trained machine learning models. We will dissect various attack vectors across different adversarial settings and their potential to compromise data. Additionally, we will discuss strategies to mitigate these attacks and showcase the effectiveness of such techniques.
Finally, we will touch upon the legal aspects and the importance in bridging between legal and technical definitions of risk.

Juulia Suvilehto

Senior data scientist @ Sahlgrenska University Hospital

Characterizing High-Needs High-Cost Patients with Segmentation

Healthcare utilization is extremely unevenly distributed in the population. Earlier research suggests that up to 50% of resources of a given healthcare system are used by only 5% of patients. This group is often called high-needs high-cost patients.  While the existence of this high-needs high-cost patient group is well established, the group has not been properly characterized. We know that these patients are likely to have multiple chronic health conditions and that they need a lot of care to manage their health, which cannot be properly modeled using industry standard methods of inspecting patient groups one diagnosis and one visit at a time. In this talk,  Juulia presents work done at Västra Götalandsregionen to detect and characterize this patient group using unsupervised learning. She will share some early findings and lessons learned from introducing advanced analytics to an organization more familiar with reporting and classical statistics.

Fabian Peng Kärrholm

Senior Data Scientist @ Volvo Cars

Scenario Extraction in the Real World

To develop a self-driving vehicle, there’s a large need for real world scenarios to be used for validation and verification of the autonomous drive function. Real world scenarios need to be collected by vehicles driving in the intended operational domain of the function. This work focuses on evaluating the collected data in the form of time-series from those vehicles, and classifying it into different states, events and scenarios. The found items in the data can then be used to conduct simulations of the AD function, risk analysis of intended launch areas and traffic behavior assessment. The presentation will cover different methods to analyze and structure the data, and ways to draw conclusions about traffic behaviour based on it.

Olof Harrysson

Machine Learning Engineer @ Modulai

GenAI: Summary of Breakthroughs in Image Generation

This presentation offers a comprehensive overview of the latest breakthroughs in generative AI, specifically within the realm of image generation. We highlight the capabilities of influential models such as Stable Diffusion and MidJourney, accompanied by a showcase of stunning and creative images they have produced.
We address the significant contributions of LORA in the development of custom models. These models are specifically designed to generate images with particular styles or concepts, marking a significant leap in personalized content creation.
Further, we explore innovative features like ControlNet, which gives users the ability to shape the structure of generated images, and IPAdapter, which introduces the concept of image-based prompts for generation.
Our focus then shifts to the dynamic open-source community, a treasure trove of diverse tools, techniques, and models that continuously fuel progress in this field. Among these innovations, we highlight ComfyUI, a tool that has been aptly likened to the “Photoshop” of image generation, due to its versatility and powerful features.
Our conversation extends beyond the technical intricacies to include the broader business implications, illustrating how these advancements are revolutionizing industries.
To conclude our presentation, we will peer into the future of image generation, anticipating further innovations and exploring the evolving landscape of this fascinating field.

Robin Johansson

COO @ Optima Planta

The Role of Data and AI in Farming

The farming industry is facing many challenges, both in terms of economics and in terms of sustainability. Financially, it is a low margin industry with rising costs. Environmentally, we see a growing population which is problematic when combined with limited resources, usage of 70% of the freshwater, and a third of the arable land having been destroyed in the last 40 years.

So what role does AI play in the future of farming?

Robin will talk about the problems in the farming industry, the potential applications and impact of data and AI, what is already done today, and what and what is needed to see the results of AI within farming.

Pier Luigi Dovesi

Senior AI Engineer @ Silo AI

Adapting Like Humans: Embodied AI Beyond Datasets and Domains

The traditional approaches of training artificial intelligence on fixed datasets within defined domains have significantly advanced the field. However, the dynamic and evolving nature of real-world scenarios requires AI systems to break free from these boundaries, mirroring human adaptability.

This talk unveils the journey towards embodied AI systems, capable of continuously learning across diverse contexts, much like humans. Inspired by recent works on real-time adaptation, we outline how these methodologies serve as stepping stones towards embodied AI.

Embodiment in AI is important because it captures the agent’s ability to perceive, interact, and symbiotically evolve with the environment. We will show how transitioning from static datasets and domain-dependent AI to a more dynamic, embodied AI, can enhance the robustness and applicability of AI systems in the real world. This requires AI to have the ability to adapt in an online learning fashion, responding to the ever-changing conditions it encounters.

Our research is now going beyond academia becoming foundational tools for AI innovators. One example is Rain AI, startup backed by Sam Altman and closely tied to OpenAI, working on neuromorphic computing. Rain is using our work on real-time adaptation as a key component of their technology stack.

Andrea Krogdal

Data Scientist @ Eghed

Key Insights (and Frustrations) Alongside the Technical Journey from a Fluffy AI Vision to a Complete ML-Product, Viewed from a Gritty Data Scientist Perspective

How do you go from a fluffy AI idea to a concrete ML product? The ambition is to clarify this fluffiness based on experiences from a completed project at Stena Stål with an end-to-end implementation predicting the volatile prices of steel. Focusing on key learnings from a data scientist’s perspective, we present an effective constellation of well-established technical solutions alongside emerging ML tools, such as MLflow, to produce a stable, repeatable and quality-assured machine learning workflow. 

 After covering the technical part, we will also try to give input on how to solve the really tricky part – how to intrigue stakeholders by mastering the art of explaining the concept of ML and how to stop wasting people’s skills and instead let them do the job they actually thrive in. Because then the ML-stars are aligned!