AI Sweden has a mission to ignite and accelerate AI in industry and society through collaboration across a wide range of partners. Therefore, we have defined a number of strategic areas that would not only solve key issues for accelerating and industrializing AI in Sweden, but also provide platforms in which several partners, from academia, public sector and industry can collaborate.
AI has enormous potential to catalyse new products and services and make existing processes more efficient. However, there are still huge challenges involved in building robust, large AI applications efficiently. To address these, we focus on the operationalization of AI - making sure that AI methods can be used at scale in critical systems and real-world scenarios, and AI systems can be developed efficiently and in a principled manner.
We will prioritize areas that can contribute to Swedish AI sovereignty and independence, i.e. breaking dependencies on external technologies and platforms critical for Swedish AI utilization – not in the sense that we should not leverage the best technologies and platforms available, but in that we are prepared to move and rebuild if necessary.
In depth: Strategic areas
Take part of a webinar where Daniel Gillblad, Co-Director Scientific Vision introduces our strategic areas.
Strategic areas in depth
Re-usable models for Sweden
While designing and implementing specialized AI and ML models for specific tasks and datasets is and will be commonplace, it is becoming more and more common to leverage existing knowledge bases, components and pre-trained models that are then specialized for the task at hand, often with much less data. Within natural language processing, models such as BERT, GPT and their descendants are trained on large datasets and then specialised for specific tasks in specific contexts and domains, and within computer vision it is often possible to start from a pre-trained model for a wide variety of tasks. However, building these general models can often require large amounts of data and computational resources.
An important area for Sweden to accelerate the use of AI is to make sure these kinds of general models and components, specifically ones addressing areas where there is less interest internationally such as Swedish language models and Swedish speech-to-text components, are broadly available and usable for both the public and private sector.
Privacy preserving AI
When and how to share data and information is a central question for all companies and government agencies working with data driven methods. While there might be immense value in data outside of an organization that controls it, dissemination is often very difficult due to privacy concerns (the data might contain e.g. personal information) or concerns about how much that data might divulge about the business and what competitive advantage it provides.
To gain full value of AI and data driven methods in Sweden, we need to find ways to work with sensitive data in a Swedish context. This means not only developing and understanding methods for privacy-preserving data- and model sharing, but also legal aspects and best practises when collaborating and implementing AI solutions. As such, privacy preserving AI is a key area for accelerating AI in Sweden on a systemic scale.
Decentralized AI – Edge and federated learning
While centralized AI systems with access to all data and information in the cloud or in a device are easier to engineer and implement, decentralised systems are becoming more and more important. Due to data privacy restrictions and limited bandwidth, systems may have to learn locally using e.g. federated learning, or act together as a multi agent system.
As data sharing and system autonomy are key limiting factors for a large number of potential applications in e.g. IoT or on sensitive personal data in government agencies or healthcare, decentralized AI will play a critical role in the use of AI in society. Here, we believe there is a very large potential for collaboration in Sweden.
AI systems, platforms and operations
In practice, core AI algorithms and methods are usually only a small part of a complete AI or ML application. Data management, resource management, and operations are often larger parts demanding far more effort in development and maintenance.
To be able to deploy real AI applications fast and efficiently, the maturity of and knowledge around systems and platforms supporting the development and operations of real-world AI systems are essential. For Sweden to be able to use AI at scale and to be relatively independent in terms of systems and platforms, we need systems research and development in this area. As many organisations share the same challenges, collaboration around AI platforms, operations of AI systems, and processes for building data driven systems should be encouraged.
Complex systems management: ML and OR
A large part of the potential value of AI lies in making operations and processes more effective, such as better resource use in cloud and telecom systems, more efficient supply chains and production, and better diagnostics and resource management in healthcare. Automation of these systems typically rely on both data driven methods for example for anomaly detection, diagnostics, root cause analysis, and demand forecasting, and capacity and resource planning and scheduling.
The intersection of machine learning and operations research has the potential to make many industries and organisations dramatically more efficient, not least medium-sized companies. As this type of application is central in Sweden, both industry and public sector, we need coordinated initiatives to build and share competence, knowledge and platforms.
While complex and data-driven systems may perform very well on average and in situations similar to those that have been previously encountered, they may behave worse or unpredictably when encountering unknown situations and data - sometimes even when data is very similar to previously observed examples. This is acceptable in situations where the cost of making a mistake, a wrong classification or prediction is low, but often not in e.g. industrial or medical settings. Here, we often need guarantees, predictable failure modes, and robustness against noise, anomalies, unknown situations and malicious manipulation. For industrialisation of AI, we need verifiable robustness in complex AI systems. As a shared challenge across sectors, this could be a useful collaboration context for many organisations.
Decision making under uncertainty
When humans and machines make decisions based on data, we typically do this under a degree of uncertainty. We usually cannot be exactly sure if a CT-scan contains indications of a complication, or if the image from a camera in an autonomous vehicle contains children playing, and this uncertainty must influence our decisions.
In general it is a challenge to quantify uncertainty in AI models, both that stemming from noisy data and that stemming from uncertainties in our model, but it is often critical to make the right decisions and to find the best solutions or plans, especially if the cost of making a mistake is high. This is often the case in real industrial scenarios, healthcare, and public decision making, and is of critical importance to Sweden given the areas where we see a large potential of applying AI. As approaches and solutions might be similar between non-competing sectors, there should be a large potential for collaboration in Sweden.