AI Safety

Understanding the nature of artificial intelligence is one of the grand challenges of our times. This challenge, together with safe deployment of AI systems, is at the core of our research on the mathematical foundations of machine learning. We engineer learning algorithms and develop data representations supporting AI interpretability, explainability, and trustworthiness. We design experiments and methods to unpack black boxes, including generative AI and large language models (LLMs). We care about AI ethics, AI evolution, and its security.

AI in Deep Brain Stimulation (DBS) and Tractography, our moonshot! During the deep brain stimulation surgery for Parkinson’s disease, the goal is to place the permanent stimulating electrode into an area of the brain that becomes pathologically hyperactive. This area is small and deep within the brain, so the main challenge is its precise localization. At NASK, we have developed a technology that calculates the localization and significantly (two or three times!) reduces the open-brain surgery time.

But that is not all. Neurosurgeries impose risks given the delicate nature of the brain tissue. Preoperative planning is critical. To address that challenge, we have also developed trustworthy AI-based methods identifying the location of eloquent brain regions and the nerve pathways leading to them.

#AI in Deep Brain Stimulation

 

In the field of natural language processing (#NLProc), we are looking for new, explainable representations of language data and investigating the relationships between grammar and semantics, particularly in tasks involving identifying and classifying atypical, potentially harmful content (such as hate speech, pornography, and misinformation). We prioritize models tailored to the problem rather than just trendy ones, resulting from a deep understanding of the data, its structures, or textual style. We are looking for the sentiment and a range of styles that can be encapsulated within algorithmic frameworks. We summarize, analyze, and anonymize. We explore internet content, legal documents, sales language, ChatGPT style, and literary parodies. We track and describe patterns—as both a classification tool and a linguistic object.

#StyloMetrix.

 

Biometrics and computer vision have a long history with us, including technologies for iris vitality identification. We design biometric authentication systems and address a wide scope of biometrics security. We study and develop deep learning methods, including those based on the transformer and generative models, signal and image processing algorithms that support deepfake detection, monitoring systems, and mobile devices.

 

Data science means machine learning algorithms, global optimization, and heuristic methods. We create decision support systems, data mining technologies, and data-driven forecasting algorithms. We study social networks focusing on cyber security issues, disinformation, and illegal content propagation.

 

We are working on:

• Detection and analysis of harmful online content
• Processing and classification of legal and formal documents
• Mathematical foundations of artificial intelligence and data analysis
• Analysis of machine learning and machine intelligence processes
• Learning and data representation engineering
• Safety, hygiene and transparency of AI (OSH-AI)
• Ethics and evolution of AI
• Multimedia data analysis
• Applications of AI in cybersecurity, medicine, biometrics, digital humanities and social research
• Digital humanities