At the Estonian Centre of Excellence in AI, our research is divided into two main categories: foundational research and applied research. We have four focus areas dedicated to the foundations of AI and five focus areas centered on the applications of AI. Our mission is to create methods for building trustworthy AI systems based on foundation models, and to ensure these serve the best interests of society.

Foundation models are powerful AI systems built using deep learning and neural networks, capable of performing a wide range of tasks with a single, unified framework. An example of a foundation model is a large language model, which can understand and generate human language. However, to address more dynamic and complex challenges, these models need to be integrated with other software components. Additionally, we aim to extend the knowledge and capabilities of foundation models to specialized contexts, such as the Estonian language, culture, and society. Our research seeks to lead the way in building and using trustworthy AI, pushing the boundaries of what foundation models can achieve and how they can be applied effectively and securely in real-world scenarios.

Below, we describe our focus areas in detail.

Hybrid AI Pipelines

In AI systems, a single foundation model can handle many tasks, but for more complex problems, it is essential to combine these models with other software tools. For example, making decisions based on data in a dynamic environment requires using databases and planners in addition to the foundation models. Our research in Hybrid AI Pipelines explores how to create well-coordinated software systems that combine foundation models, dynamic databases, and specialized algorithms. The goal is to develop more effective and trustworthy systems for analysis, prediction, and recommendations.

Adaptation of Foundation Models

Foundation models have limitations, especially when applied to specialized areas not well represented in their training data, such as medical texts or languages with fewer resources. Our work in Adaptation of Foundation Models investigates how to best use additional data that is specific to certain domains or languages. This data can be sensitive or confidential, so we also look at how to handle it securely. In particular, we focus on training models that are more knowledgeable about the Estonian language, culture, and society. Our aim is to enhance foundation models to perform better in these specialized contexts and to be more relevant to local needs and characteristics.

Safeguards and Trust in AI

While foundation models are powerful, they also have weaknesses in terms of robustness, safety, and trust. To mitigate these risks, a comprehensive approach is necessary, covering everything from the architecture of the AI system to the methods used for deploying it. Our research in Safeguards and Trust in AI focuses on developing control mechanisms and safeguards to ensure that AI systems are trustworthy, robust, and accurate. This includes implementing measures to prevent errors and biases, and to ensure that the systems perform reliably in real-world situations.

Privacy and Security in AI

Protecting sensitive or confidential data in AI systems is a critical challenge, especially as these systems become more complex. In our Privacy and Security in AI focus area, we study how to develop and implement security frameworks that protect sensitive information throughout the entire AI system. This includes ensuring compliance with data protection and AI regulations. Our goal is to create AI systems that not only secure private data but also adhere to legal and ethical standards.

AI for E-governance

E-governance involves the use of digital tools and systems to improve the efficiency and accessibility of government services. Our research in AI for E-governance focuses on developing foundation model-based AI systems that can analyze and predict the needs of citizens. By understanding these needs, we aim to create more effective and accessible services. Additionally, our work supports civil servants by providing decision support systems that help them make informed and timely decisions. The goal is to enhance the interaction between citizens and government through intelligent, responsive systems.

AI for Healthcare

Healthcare generates vast amounts of data, particularly through electronic health records. Our research in AI for Healthcare aims to develop foundation model-based AI systems that can analyze this data to improve health outcomes. This includes modeling health trajectories, predicting risks, summarizing data, extracting important facts, and detecting anomalies. By leveraging AI, we strive to enhance patient care, support medical professionals in their decision-making, and ultimately improve public health.

AI for Business Processes

Businesses can greatly benefit from AI to optimize their operations. Our focus in AI for Business Processes is on creating foundation model-based AI systems that can identify causal links between human actions and business outcomes. By understanding these links, AI systems can recommend subsequent actions or adjustments to optimize processes. This helps businesses become more efficient, reduce costs, and improve overall performance.

AI for Cybersecurity

Cybersecurity is critical in protecting information and infrastructure from cyber threats. Our research in AI for Cybersecurity focuses on enhancing security through foundation model-based AI systems. These systems can automate log management, correlate data across cyber and physical spaces, and provide advanced analytics for security monitoring, incident handling, and digital forensics. By improving these areas, we aim to create more robust and resilient cybersecurity measures that can protect critical systems from emerging threats.

AI for education

Education is a crucial area where AI can make a significant impact. Our research in AI for Education aims to develop and implement a range of AI tools that enhance learning experiences and outcomes. A key focus is the creation of AI assistants that support self-regulated learning, helping students to manage and optimize their own educational journeys. These tools will be designed based on insights from cognitive science and neuroscience to ensure they effectively aid comprehension and retention. Furthermore, we are committed to ensuring these tools are reliable and trustworthy for classroom use. Our approach emphasizes collaboration with educators to co-create these tools, ensuring they meet real-world educational needs, and includes rigorous validation with students to ensure effectiveness and usability. By advancing AI in education, we aim to create innovative solutions that support and enrich the educational experiences of students and teachers alike, fostering a more effective and engaging learning environment.