https://seminars.cs.fel.cvut.cz/ Recent content in Home on Computer Science Seminars Hugo en <a href="https://creativecommons.org/licenses/by-nc/4.0/" target="_blank" rel="noopener">CC BY-NC 4.0</a> Wed, 24 Jun 2026 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2026-06-24-kubikova-adela/ Wed, 24 Jun 2026 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2026-06-24-kubikova-adela/ https://seminars.cs.fel.cvut.cz/2026-06-10-rada-jakub/ Wed, 10 Jun 2026 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2026-06-10-rada-jakub/ https://seminars.cs.fel.cvut.cz/2026-05-27-hadzic-armin/ Wed, 27 May 2026 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2026-05-27-hadzic-armin/ Deep generative models are everywhere now. They generate photorealistic images, videos, music. Models like diffusion models, GANs, and VAEs are incredibly expressive — but there is a funny paradox hiding underneath all this progress. Even though these models are probabilistic by design, actually using them as probabilistic models is often computationally hopeless. Want to compute exact likelihoods? Condition on missing data? Perform exact inference? Suddenly the results disappear behind slow computations. https://seminars.cs.fel.cvut.cz/2026-05-19-cap-michal/ Tue, 19 May 2026 14:00:00 +0200 https://seminars.cs.fel.cvut.cz/2026-05-19-cap-michal/ Outdoor trailer logistics in industrial campuses and distribution hubs remain almost entirely manual. At ISEE, Michal and his team built the world&rsquo;s first driverless trailer yard, where self-driving trucks move trailers in mixed traffic on private roads. In this talk, he will explain how autonomous driving on private industrial property differs from the public-road use case, why existing self-driving stacks fall short, and how ISEE&rsquo;s autonomy stack addresses these challenges, including their use of reinforcement learning for fluid navigation in cluttered yards. https://seminars.cs.fel.cvut.cz/2026-05-13-milec-david/ Wed, 13 May 2026 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2026-05-13-milec-david/ In two-player zero-sum games, game theory provides an optimal solution that guarantees a minimum payoff regardless of the opponent’s actions - the so-called value of the game. This approach performs well against near-optimal opponents. In practice, however, we often face opponents who deviate significantly from optimal play, such as human players, and who can be exploited for substantially higher gains. Yet, the optimal strategy typically achieves only a small margin of victory in such cases. https://seminars.cs.fel.cvut.cz/2026-04-29-bajgar-ondrej/ Wed, 29 Apr 2026 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2026-04-29-bajgar-ondrej/ For AI systems to act autonomously and safely, they need a clear specification of what outcomes are desirable—but writing down such a specification by hand is notoriously difficult. Inverse reinforcement learning (IRL) offers an alternative: inferring what humans value by observing how they behave. However, any finite set of demonstrations leaves the underlying objective underdetermined. Bayesian IRL addresses this by maintaining a distribution over plausible reward functions rather than committing to a single best guess, enabling both robust decision-making under uncertainty and active collection of additional human input where it is most needed. https://seminars.cs.fel.cvut.cz/2026-04-22-sladic-muris/ Wed, 22 Apr 2026 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2026-04-22-sladic-muris/ Cyber deception has traditionally focused on deploying decoys to achieve early and high-confidence detection of adversaries. However, the design of effective deception strategies remains largely manual, ad hoc, resource-consuming, and yet offers little insight into the attackers&rsquo; actions. In this work, we explore the use of large language models (LLMs) to automate the design of tailored deception strategies, shifting the focus from static deployment to engaging, and context-aware deception engineering. https://seminars.cs.fel.cvut.cz/2026-04-16-krutsky-martin/ Thu, 16 Apr 2026 16:00:00 +0200 https://seminars.cs.fel.cvut.cz/2026-04-16-krutsky-martin/ Mechanistic interpretability (MI) has become the frontier technique for explaining Transformer models, however, it relies on computationally expensive, trial-and-error experiments. In parallel, neurosymbolic AI (NeSy) has been successful in incorporating formal logic into neural networks, but often suffers with scalability issues. In this seminar, I will discuss my early attempts at a synthesis of NeSy and MI, making use of their respective strengths. The seminar will serve both as an update on my PhD studies, a preparation for the professional debate, and an opportunity for receiving feedback on my work in progress. https://seminars.cs.fel.cvut.cz/2026-04-16-responsible-ai-coffee-chat/ Thu, 16 Apr 2026 14:30:00 +0200 https://seminars.cs.fel.cvut.cz/2026-04-16-responsible-ai-coffee-chat/ Join us for a Responsible AI chat covering early 2026 developments. Alongside recent reports showing a spike in requirements for human validation for AI agents, we will discuss academic pushback against AI-generated peer reviews at ICML, fresh research on the dangers of “sycophantic” chatbots and latent model biases, and how the Pentagon’s disputes over AI safeguards and Ukraine’s battlefield data sharing reflect wider questions around militarization. We’ll also explore evaluation tools such as the “Bullshit LLM Benchmark. https://seminars.cs.fel.cvut.cz/2026-04-08-ryzak-david/ Wed, 08 Apr 2026 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2026-04-08-ryzak-david/ In this talk we introduce a definition of coalitional trust which we derive from the principle that external opinion for groups should matter as it matters for individuals. We model the game for a weighted graph representing the local trust relations. Framed as a cooperative game, the model admits tractable computation of the Shapley value for specific choices of the game. We also discuss experimental results for a P2P network which raise new questions about how to design the game to prevent some specific types of behavior. https://seminars.cs.fel.cvut.cz/2026-03-25-garcia-sebastian/ Wed, 25 Mar 2026 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2026-03-25-garcia-sebastian/ Modern AI and cybersecurity systems must make decisions from noisy, incomplete, and often adversarial signals without overreacting and harming normal operations. This seminar explores the biological immune system not as a metaphor, but as a source of design principles for intelligent defense: distributed sensing, context propagation, multi-signal activation, adaptive aggressiveness, memory, and regulated shutdown. It also examines why many earlier cyber-immune approaches failed by copying isolated mechanisms instead of the larger architecture, and shows how these ideas were already translated into a real implementation. https://seminars.cs.fel.cvut.cz/2026-03-11-karapetyan-ruben/ Wed, 11 Mar 2026 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2026-03-11-karapetyan-ruben/ Many optimization problems can be solved indirectly by transforming them into a different mathematical representation. In polynomial optimization, one common route is the moment-SOS hierarchy, which returns a matrix containing statistical information about the solution together with the optimal value instead of returning the minimizing point itself. This talk focuses on how to recover the actual minimizer from that moment information when the available data is infinite-dimensional. It introduces the new challenges that appear in that setting and asks whether the underlying optimizer can still be recovered from such finite moment data. https://seminars.cs.fel.cvut.cz/2026-02-25-hulchuk-seva/ Wed, 25 Feb 2026 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2026-02-25-hulchuk-seva/ Tracking robot position is straightforward when GPS is available, but autonomous navigation becomes harder underground or in other GPS-denied settings. A 3D LiDAR scans the surrounding world and can localize directly against a previously built 3D model, but this approach depends heavily on the environment&rsquo;s geometry. This seminar looks at the harder cases, such as tunnels, open fields, and mountain meadows with fewer geometric features. It covers how 3D LiDAR localization works, where it struggles, how multiple sensors can be fused to handle localization failure, and how LiDAR-based scan matching connects to ongoing work on LiLi and Lie-theory descriptions. https://seminars.cs.fel.cvut.cz/2026-02-11-straka-matej/ Wed, 11 Feb 2026 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2026-02-11-straka-matej/ SAM 2 achieves strong video object segmentation but relies on rigid memory update rules, while recent heuristic improvements struggle to generalize. We present SAM2RL, which formulates memory management as an MDP and learns update policies via reinforcement learning. A lightweight PPO-trained agent using only occlusion scores and temporal distances achieves state-of-the-art results across multiple VOS benchmarks. We also show that many reported heuristic gains fall within the variance of random memory selection, questioning current evaluation practices. https://seminars.cs.fel.cvut.cz/2026-01-28-rigaki-maria/ Wed, 28 Jan 2026 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2026-01-28-rigaki-maria/ Cybersecurity environments often employ vectorized representations of the state and action spaces to leverage existing reinforcement learning (RL) libraries. However, these representations suffer from two major flaws: either they leak privileged information to agents, or they fail to scale as the network grows. Contrary to this approach, NetSecGame defines states through a flexible representation of sets of hosts, services, data, etc., allowing the agent to specify how to use the available information. https://seminars.cs.fel.cvut.cz/2026-01-14-maskova-michaela/ Wed, 14 Jan 2026 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2026-01-14-maskova-michaela/ In financial transaction monitoring and anti-money laundering, improvements in machine learning-based detection are fundamentally limited: relevant information required for accurate decision-making is often not present in transactional data alone and only becomes available through human investigation. At the same time, regulatory constraints require that final judgments remain with human analysts. In our work, we design a transaction monitoring framework that explicitly incorporates human-AI interaction to support analyst decision-making. We model the sequential investigative process of analysts using signals extracted from historical investigation reports. https://seminars.cs.fel.cvut.cz/2025-12-17-drabent-karolina/ Wed, 17 Dec 2025 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2025-12-17-drabent-karolina/ In large games, a common practice is to approximate an opponent using a small set of representative strategies. This idea works well in practice, but the portfolios themselves are usually built using heuristics with little theoretical backing. We study this problem formally in two-player zero-sum games. I will show that finding the best possible portfolio is NP-hard, and that several intuitive approaches, such as using Nash equilibrium supports, can fail badly. https://seminars.cs.fel.cvut.cz/2025-11-19-cuadrado-avila-nicolas-mauricio/ Wed, 19 Nov 2025 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2025-11-19-cuadrado-avila-nicolas-mauricio/ Long-term power system planning must operate under deep uncertainty, yet traditional finite-scenario methods miss critical volatility. We model high-dimensional uncertainties with tractable probabilistic models, specifically sum-product networks, to enable exact inference and enforce reliability through chance-constrained planning. https://seminars.cs.fel.cvut.cz/2025-07-30-discussion/ Wed, 30 Jul 2025 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2025-07-30-discussion/ Check out the notes from the seminar. FEL account login is required. There are so many AI tools now, so this session focused on how to use them effectively in research work. The examples ranged from programming assistants such as Copilot or Cursor to literature review workflows and writing support with ChatGPT or Gemini. The session was designed as a discussion preceded by a few short talks, with shared notes gathered afterwards so the group would leave with something concrete. https://seminars.cs.fel.cvut.cz/2025-04-23-valeros-veronica-garcia-sebastian/ Wed, 23 Apr 2025 14:30:00 +0200 https://seminars.cs.fel.cvut.cz/2025-04-23-valeros-veronica-garcia-sebastian/ In this seminar, we share how we are changing cybersecurity education with our Introduction to Security class. We will share what are the pillars of our class, how we organize the content and the labs to encourage students to practice, break things, and learn. We will share our experience of opening up the class as a Massive Open Online Course and how we taught 1600 students in Winter 24/25. We will also introduce our StratoCyberLab, a cyber range that we developed for students to practice their cybersecurity skills at home in a safe environment. https://seminars.cs.fel.cvut.cz/2025-03-26-kubicek-ondrej/ Wed, 26 Mar 2025 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2025-03-26-kubicek-ondrej/ The combination of reinforcement learning and search has been crucial in training superhuman agents in two-player zero-sum games with imperfect information. However, sound search algorithms in those games have limited scalability, as Texas Hold&rsquo;em Poker is the largest game where those algorithms were successfully applied. This talk covered a method for scaling those algorithms beyond poker, along with the high-level design ideas, the limitations of the resulting algorithm, and the open questions that remain. https://seminars.cs.fel.cvut.cz/2025-02-26-nemecek-jiri/ Wed, 26 Feb 2025 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2025-02-26-nemecek-jiri/ Machine learning often gives us good models for specific tasks, but practical decision problems can require optimization under multiple, stricter, or difficult-to-define criteria. That can mean optimizing over one or more trained models directly. This seminar used counterfactual explanations as the running example, focusing on plausibility and showing how and why it can make sense to optimize over a neural network and a sum-product network at the same time. https://seminars.cs.fel.cvut.cz/2025-01-29-janota-jirka/ Wed, 29 Jan 2025 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2025-01-29-janota-jirka/ Everybody asks where the honeybees are, but nobody asks how they are. This talk briefly presented an approach to mapping the honeybee comb and why that mapping is useful in practice. From there, it moved into the computer vision challenges of working inside a living honeybee colony, including what methods worked, what did not, and where the audience should have pushed back. https://seminars.cs.fel.cvut.cz/2024-12-18-fiedler-david/ Wed, 18 Dec 2024 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2024-12-18-fiedler-david/ What to expect: a brief presentation about the most common mistakes in experiment organization and how to avoid them, followed by a discussion of good practices and tools for experiment organization and automation. The target audience was anyone already running computational experiments or planning to, whether locally or on the RCI. No prior requirements were needed. https://seminars.cs.fel.cvut.cz/2024-11-27-kovarik-vojta/ Wed, 27 Nov 2024 13:30:00 +0200 https://seminars.cs.fel.cvut.cz/2024-11-27-kovarik-vojta/ The field of Cooperative AI is concerned with ensuring that interactions between AIs, or between humans and AIs, lead to beneficial outcomes. This seminar argued that while game theory is our main tool for modelling strategic interactions, it was designed primarily with human agents in mind and does not fully fit AI settings. The second half turned to situations where agents can effectively read each other&rsquo;s minds and what that does to the assumptions behind standard game-theoretic analysis. https://seminars.cs.fel.cvut.cz/submit/ Mon, 01 Jan 0001 00:00:00 +0000 https://seminars.cs.fel.cvut.cz/submit/ Would you like to present? We are always looking for exciting new research, practical insights, and methodological discussions. Whether you are a PhD student looking to practice a presentation, or a researcher wanting to share your latest findings, we would love to feature your talk! Feel free to add any thoughts or topics; the overall goal is an open forum to exchange ideas about what we do! 🎤 Sign up to present via our Google Form