Constraint-Based Solution Methods for Automated Production Scheduling

Modern-day factories often use a complex and highly automated process to manufacture large quantities of products. Thus, finding an efficient production schedule becomes challenging, as usually, many complex constraints impose restrictions on feasible schedules. To capture the requirements of production scheduling problems from various domains, high-level constraint modeling languages, which can be used together with automated constrained optimization solving technology, have been proposed in the past. However, to handle large-scale scheduling problems efficiently, novel modeling techniques and intelligent algorithms are often needed to tackle the complex scheduling constraints in real-life environments.

This talk will give an overview of our work with constraint-based solution approaches for several real-life scheduling problems from the industry. Thereby, we will discuss advanced constraint modeling strategies and state-of-the-art solution algorithms.

Understanding Large Language Models and the Path to AGI

Neural Networks are often perceived by the general population as a form of magic, but at their core, they are essentially a structured sequence of mathematical transformations mapping an input tensor space to an output tensor space. Large Language Models (LLMs), such as ChatGPT, operate through a series of tensor algebra operations, leveraging vast amounts of data and computation. The true "magic" emerges not from individual calculations but from scaling-when models grow larger, they exhibit emergent properties that were not explicitly programmed.

This talk explores the implications of scale in AI, drawing lessons from nature. Evolution did not grant humans 86 billion neurons and 100 trillion synaptic connections by accident; nature is economical, and the complexity of human intelligence is deeply tied to its capacity. The human brain's encephalization quotient-the ratio of brain mass to body size-exceeds that of any other primate, highlighting the importance of scale in biological intelligence.

A central question arises: are human intelligence and consciousness Turing-computable? If intelligence is simply the product of sufficient capacity and complexity, then in principle, AI models, when scaled, should be able to achieve human-level Artificial General Intelligence (AGI). But does the nature of intelligence go beyond computation? What is the Kolmogorov complexity of human intelligence? The Chinese Room argument, proposed by philosopher John Searle, challenges the idea that syntactic manipulation alone is sufficient for true understanding. Meanwhile, philosopher and cognitive scientist Daniel Dennett's theories on consciousness suggest that intelligence is just an emergent property of information processing, much like what we observe in modern AI models.

This talk will critically examine these perspectives, discussing whether AI is on the trajectory to achieving human-like cognition or if there are fundamental barriers that limit computational models from replicating consciousness. Ultimately, we will explore whether the rapid scaling of AI is bringing us closer to AGI or revealing the limits of algorithmic intelligence. Is the human brain super-Turing powerful?

How to deal with failure in GACR project submissions?

What are the possible reasons for the failure of project proposals? What can we do about it? Come and hear the experiences of GACR panel members and discuss ways to improve the success rate of our proposals. We will present our proposals and would be happy for your feedback and any suggestions you may have. Since science funding is a critical aspect of a healthy faculty research environment, we need to confront the ongoing challenges and take the necessary actions. By your participation in this discussion, you can directly participate in shaping these measures.

Multimodal TimeViz: Visualization and Multimodal Exploration of Time-Oriented Data

In this talk, I will explore the complexities and nuances of visualizing time-oriented data. Time, as a unique dimension, possesses a rich semantic structure that extends beyond a simple linear progression of values. I will demonstrate why effectively modeling this structure is essential for capturing temporal irregularities and cyclic patterns. Through practical examples, I will illustrate how these temporal characteristics significantly impact visualization design and interpretation. The talk further explores emerging research directions, including visualization literacy, audio-visual analytics, situated visualization, and knowledge-assisted visual analytics.

A data-driven approach in medical education and healthcare in practice

This lecture provides a comprehensive view of dependent outputs and activities in academic and governmental sectors that concentrate on health and medical education data collection, processing, analysis and visualisation, including the correct interpretation. Recognising the importance of correctly managing, identifying, and disseminating accurate information and insights from the vast array of data of differing quality is essential.

The lecture includes case studies that exemplify new findings and good practices in technology-enhanced learning and health informatics based on experiences from research and development projects, stakeholders' opinions, and community requirements. There are proven methodological and technological backgrounds, emphasised in this lecture, that cover a complete and robust framework for two unique and multidisciplinary domains of human interests and understanding: first, medical and healthcare education supported by the MEFANET network; second, health information and statistics describing mainly data from the National Health Information System. These topics are set in the context of achieved results applied in daily practice, sharing know-how, student involvement and community building. Generally, this research field is multidisciplinary and summarises the results achieved, combining the fields of informatics, medical education, pedagogy, data analysis, and health literacy.

What is my quantum program doing?

Development of quantum programs is hard due to their intricate structure and inherently probabilistic nature. Computer-aided tool support is therefore essential. Computer-based reasoning over quantum programs is, however, also challenging due to the exponential size of the program's state. In this talk, I will present a recent framework for automated formal verification of quantum programs that uses automata to represent complex sets of quantum states compactly.

CoFI break with management

The Lab Fest, held as part of the Informatics Colloquium, aims to introduce the activities of research groups at the Faculty to the academic community, specifically their staffing, areas of interest, involvement in grants, ongoing collaborations at the University, in the Czech Republic or internationally; future directions and current research or development results.

Jan Křetínský: Learning in Verification group (LiVe Lab)

Michal Kozubek: Centre for Biomedical Image Analysis (CBIA)

Jumping from Academia to Startup-Land: One Man's Story of Transferable Skills

Leaving academia for an academic spinout can, at first, feel like starting any other research project. But at some point, the differences start to become clear. While the technical challenges may be familiar, the priorities, constraints, and measures of success are often quite different. The objectives of a research project don't always align with those of a company, the motivations of funders shift from knowledge generation to commercial impact, and managing a team in a startup comes with its own unique dynamics.

In this talk, I'll share my experience navigating this transition: the lessons I learned, the mistakes I and others made along the way, and the aspects of academic training that turned out to be useful. Whether you're considering a similar move, want to make your teaching more relevant to students going to the industry, or are just curious about how research skills translate to startups, I hope my talk will give you some useful insights.

Kristýna Pekárková: Twin-width of Graphs on Surfaces

Twin-width is a structural graph parameter introduced in 2020 by Bonnet, Kim, Thomassé and Watrigant. Graphs of bounded twin-width generalize a wide range of graph classes, both sparse and dense, and the parameter has turned out to be an important tool in many areas, including structural and algorithmic graph theory, combinatorics, and model theory. Although it is known that many graph classes have bounded twin-width, the bounds are often enormous, and determining better, explicit bounds has, therefore, been a focus of an intensive line of research. In this talk, we present a new, asymptotically optimal upper bound for twin-width of graphs embeddable on surfaces.

Ján Jančár: Peeking into the black-box: Reverse-engineering cryptographic implementations

Cryptography secures our online communications, payments and other services. We tend to rely on it without questioning its security and security of cryptographic implementations. Software implementations, especially open-source ones are auditable, although that is not a guarantee of security. However, cryptographic implementations in hardware, such as smartcards, are closed-source with extremely limited publicly available information and their details guarded under NDAs. Users thus have no way to independently verify the security claims, and are left trusting security certifications that were shown to be lacking in the past.

In this talk we will discuss several methods of "peeking into the black-box" and gaining insight into the implementation details of elliptic curve cryptography implementations. We will see how simple algebraic tricks can be used to extract closely guarded details from smartcards from major manufacturers. Finally, we will discuss what this means for the secure hardware space and what manufacturers can do to defend themselves.

What challenges do biomolecules bring to IT?

Biomolecules, such as proteins and nucleic acids, are 3D objects that are essential building blocks of living organisms. But what are they hiding inside? Various complex data structures - sets of 2D and 3D points, graphs, multigraphs, matrices, groups, etc.. The contents of these data structures make not only biological but also logical sense. Moreover, the datasets of biomolecular 3D structure are huge. On top of that, research on these data can bring breakthrough results for our life.

In my talk, I will present selected key challenges in biomolecular 3D structure research (e.g., structure visualization, prediction, similarity search, property computation) and IT ways to address them. In addition, I will show the questions that are still open and waiting for their IT solutions.

FIxbot: Practical steps towards AI-assisted teaching at our faculty, discussion

The talk will outlay initial demand, visions, and practical experience with AI-assisted teaching at our faculty, resulting in the need for teaching methodology changes as well as self-hosted models with customizable processing pipelines for better control over the interaction data for improved education.

The rise of LLM-powered programming assistants has upended conventional teaching strategies, particularly in programming courses where step-by-step skill-building has been the norm. While these models excel at solving simpler problems, they leave students struggling when faced with more complex assignments - without the foundational understanding built through progressive learning.

Our approach is based on the assumption that students - future graduates - will routinely use LLM-powered programming assistants in their work. Therefore, learning outcomes of programming courses should focus on collaborative development, in which the programmer interacts with the assistant. The burden of the developer is to accept responsibility for the result, which involves thorough verification of the specification fulfilment, testing and debugging.

We designed FIxbot - an LLM-based pipeline that inserts bugs into the code and allows for guidance in identifying these bugs if asked. We will summarize observations and lessons learned from the usage of such a pipeline for teaching PB071 (Programming in C), including technical deployment of open-weight models inside the FI infrastructure. Additionally, we will explore how students in PB173 (Domain-Specific Development) use a range of commercial and open LLM-based setups for their programming tasks.

Building FIxbot.fi.muni.cz is a work in progress and we need your insights and expertise. A large fraction of talk time will be dedicated to discussions about the additional use of FIxbot at FI, new ideas, and technical considerations.