FCAI research focuses on tackling three challenges, which are key bottleneck problems for a wider adoption of AI systems:
I DATA SCARCITY
Huge labeled data sets made the AI revolution possible. However, most of the value in big data is in the enormous number of small questions it could answer but at this resolution data becomes a scarce resource. We will widen this bottleneck and make AI applicable to a significantly wider scope of questions by developing data efficient methods, by bringing in prior knowledge in the form of models, and by enabling privacy-preserving sharing of data.
AI systems are not dependable, i.e., they are vulnerable to manipulation and information stealing, and it is not certain whether the outputs are trustworthy. We will develop the required privacy-preserving and secure AI, and methods that can take into account uncertainty in data and decision-making. We will provide new resilient deep learning approaches for the currently popular and successful deep neural networks. Societal trust stemming from dependable AI enables wide applicability in the society.
III UNDERSTANDING HUMANS
AI does not understand the user. We will provide AI with the capability to understand the user, which is a prerequisite for making AI understandable. Modeling the user and the interaction will help AI understand the user and vice versa. The outcome is AIs that are able to augment human capabilities in a multitude of tasks.
FCAI research agenda builds on our world-level expertise in machine learning, the basis for many recent breakthroughs in AI.
We run five spearhead research programs with multiple research groups. They provide the methods and results to solve the three bottleneck problems.
Agile probabilistic AI
Contributes to challenges I and II by practical probabilistic programming tools. They allow creating reliable and understandable AI models that are trained on complex small and big data. These models are easy to use and refine, supporting an agile workflow.
Contributes to challenges I and III by new methods needed for the next generation of machine learning with efficient, interpretable reasoning capability, by cross-breeding modern machine learning and simulator-based inference.
Next generation deep learning
Contributes to challenges I and II by providing the next generation of deep learning based solutions.
Coordinated by CEO Harri Valpola (Curious AI)
Privacy-preserving and secure AI
Contributes to challenges I and II with security and privacy research. We develop realistic adversary models to build effective tools and techniques that practitioners can use to build dependable AI systems.
Contributes to challenge III by improvements in acceptability and collaborative abilities of AI systems for human use. They include interactive AI that we can talk to, is capable of understanding human behavior, acts more safely, and collaborates with humans in complex tasks. New ways are explored to demonstrate solutions to AI, criticize its behavior, and steer it.
AI is a general-purpose technology and has extensive potential applications.
The most important ones in which FCAI is estimated to have competitive advantage will be actively pursued within the flagship as Applications and Cross-Breeding research programs:
Utilization of massive multi-source data in more accurate risk prediction, diagnosis and choice of treatment for an individual. AI will play a central role in this process and it is already making the first steps into the clinics in a multitude of settings ranging from the diagnostic monitoring of intensive care patients to assisting in tumor identification. Our objective is to revolutionize adoption of personalized medicine through spearheading AI research activity towards utilizing biomedical data.
AI for Health
There are now large-scale health datasets collected by hospitals and research institutes together with heterogeneous wellness data produced by wearable devices and health apps. They form an invaluable asset for both cutting-edge AI research and business opportunities in health technology. The Finnish healthcare system and research institutes are at the cutting edge of collecting and digitizing medical records, medical genetics, and in running large longitudinal cohorts where risk-individuals and their controls are followed from birth until disease onset. Our objective is to make Finland a global forerunner in data-driven, privacy preserving AI for health, wellness, and medicine.
AI and Data Science
Modern data science and modern AI both share a strong emphasis on analysis of data for intelligent decision making. The key research topics for data science include efficient and effective computational methods for analysis of various data sources, including large, heterogeneous and complex data sets, and also platform solutions taking advantage of parallelism and distribution of computation and storage (high-performance computing, grid computing, cloud computing). The participating universities have already agreed to set up a joint Helsinki Centre for Data Science, HiDATA. The objective is to create a world-class research and research-based education hub of data science in Helsinki together with HiDATA.
Coordinated by Professor Hannu Toivonen (University of Helsinki) & Professor Aristides Gionis (Aalto University)
AI and X
FCAI also has an ”AI and X”-programme, where X’s represent promising application fields such as materials physics, social sciences and economics. The objective of this programme is to maximize the impact of AI by working together with top experts of an increasing number of fields X. In the beginning, “AI and X” programme includes 30+ professors in a number of fieds, and the list will be dynamically updated. We welcome new ideas and disciplines to join our common endeavor through the “AI and X” programme.
Some examples of the 30+ professors participating in the “AI and X” programme are: