Research teams

The group focuses on modern machine vision methods based on deep learning. It works on various tasks, such as image classification, object detection, segmentation, image generation, and analysis using large vision-language models. Its research focuses on challenges related to knowledge adaptation and transfer, model robustness, control of generative systems, and combining information from different modalities. 

The team works on applying artificial intelligence methods to develop tools that support medical diagnostics based on image data (e.g., radiological, histopathological, etc.). One of the main areas of research is the development of an algorithm to support the histopathological diagnosis of lymphoma. Technical issues include tasks such as classification (e.g., cancer cell vs. healthy cell), segmentation (e.g., tumor area), and modeling of specific processes (e.g., cell differentiation).

The team develops AI algorithms for battlefield analysis and decision support in combat situations. We are building unique capabilities that use artificial intelligence algorithms to enhance the capabilities of BMS (battlefield management systems). We are working on advanced techniques of data analysis, machine learning, optimization, and game theory to use them to improve the command decision-making process at both the tactical and operational levels. Our goal is to achieve a statistically significantly higher average decision-making effectiveness compared to a human operator, while minimizing the risk of undesirable events. 

In order to build solutions to support operations in the strategic, operational and tactical areas, it is necessary to accurately map and describe the area of potential operations. The team is working on AI algorithms to analyze the battlefield and its analytical description in space from 2 to 3D. We are building unique solutions for digital twins, among others, using artificial intelligence algorithms to enhance the capabilities of battlefield management systems (BMS). Advanced data analysis and machine learning techniques, along with extensive interdisciplinary knowledge, will improve decision-making and support the armed forces in building their defense capabilities.

The team is working on modern techniques for creating 3D images, focusing particularly on NeRF and Gaussian Splatting methods. NeRF is a method that uses artificial intelligence to create realistic 3D models from photographs. It works well, but it is slow and time-consuming. Gaussian Splatting, on the other hand, is a newer method that uses distributions (called Gaussians) to describe a scene instead of neural networks. This makes it much faster and easier to edit objects. It allows for interactive changes similar to those made on classic 3D models. The team is also researching how to combine Gaussian Splatting with generative AI models to have more control over the appearance and quality of 3D scenes. 

The team conducts research related to the analysis of existing and potential future cyber threats, monitors their development trends, and develops innovative ways to counteract them. To this end, modern technologies and algorithms, including artificial intelligence methods, are used. The solutions implemented include intelligent methods for detecting vulnerabilities in software and ICT systems, detecting disinformation, and counteracting covert communication techniques.

As part of its research on graph neural networks, the team is developing neural architectures that compress graphs into numerical vectors and synthesize graphs. These methods are used in chemistry and pharmacy to model relationships between chemical compounds and generate hypothetical chemical compounds. As part of its research on multi-agent reinforcement learning, the team is developing methods of communication in a multi-agent environment. These methods are used in the optimization of control of distributed systems, such as power grids or swarms of drones. 

The team deals with large language models (LLMs), their effective training and tuning. They are investigating the possibilities of using a multi-agent approach to solving tasks, optimized agent creation, and the use of other modalities, i.e., image and sound. An important element is ensuring mechanisms for introducing multilingualism into language models. The team is also working on extracting information from multimodal data and its semantic analysis using language models.

The team conducts research related to the application of artificial intelligence in the field of mental health, in particular to understand neuroatypical experiences (including those of people with autism) based on first-person data. The team is also working on the use of large language models in therapy, studying the interactions between agents and humans from a bioethical perspective. 

The research team conducts research on methods of automatic data analysis and content classification in the context of privacy protection and regulatory compliance. In particular, we investigate the use of natural language processing (NLP) techniques and other methods to identify sensitive information and assess the risk of its unauthorized disclosure or misuse. The group's key research areas include the development of algorithms for the precise detection of protected content, semantic analysis of documents to determine their level of confidentiality, and the adaptation of classification processes to changing regulations.

The team is dedicated to using advanced machine learning techniques to model complex adaptive systems. We focus on challenges characterized by low signal-to-noise ratio, lack of stationarity, and emergent properties. In particular, we address problems with limited opportunities for controlled experimentation, such as weather forecasting, optimization and management in the energy market, inference in financial markets, and prediction of information dynamics. The team combines a research approach with practical impact on the surrounding reality.

The team focuses on theoretical analysis and investigation of the internal mechanisms of state-of-the-art generative models. We analyze how these systems represent and process information, using mechanistic interpretability techniques, among others, to discover new methods for their adaptation and control. Our goal is to translate this knowledge into resilient and transparent solutions applicable to various domains such as computer vision, sound and life sciences.

The team is dedicated to directly translating scientific research results into practical technological solutions. Our main goal is to provide dedicated artificial intelligence systems for local government units and public administration, which realistically increases the level of digitization of the country. In our daily work, we integrate technologies of natural language processing (NLP), computer vision and operations research to create tools that optimize complex decision-making processes, automate document and image analysis and measurably improve the daily functioning of public institutions. We act as a link between academic research and real business needs.