Tuesday seminar on artificial intelligence

Time: Tuesday, 10:40
Place: S8, Malá Strana.

A seminar that aims to refer on ongoing research projects of participants and related issues within the scope of artificial intelligence, theoretical computer science and mathematical logic. Presentation should be for 60 minutes and follow-up discussion is expected. If you, your student or your guess would like to give a talk, then let me know to schedule it. If you would like to receive notifications about upcoming presentations, then write me.

Students can enroll this seminar as the course NTIN091 (Seminar for MSc. and Ph.D.-students).

Summer semester 2023/24

12.3. Matej Moravčík: Bridging the Gap: Towards Unified Approach to Perfect and Imperfect Information Games (defense of a doctoral thesis)
19.3. Department meeting:
26.3. Mrinalini Subramanian:
Multi-Agent Pathfinding (MAPF) is a pathfinding problem involving a set of agents navigating from start to target nodes without collisions. This talk offers fundamental insights into MAPF, emphasizing its practical applications, key challenges, and ongoing research. Various algorithms address the coordination of movements among multiple agents. They can be categorized into two major approaches, one based on search and one based on compilation, both will be presented. The talk will conclude with discussion on future works and research questions, offering a comprehensive overview of the current state of exploration in the realm of MAPF.

Winter semester 2023/24

9.1. Alexander Ocheretyany: Neural Architecture Search: Modern approach
In recent years active research in Machine Learning has resulted in breakthroughs in computer vision, speech recognition, natural language processing, reinforcement learning and other areas. Neural Architecture Search (NAS), the process of automating the design of neural architectures for a given task, is an important step in automating machine learning that has already outperformed the best human-designed architectures on many tasks. In this talk we provide insights into NAS and modern techniques used in the area.

Summer semester 2022/23

21.2. Adam Harmanec: Tracking in live-cell microscopy super-resolution videos
Automated processing of biological data is essential for an effective and unbiased analysis. We study super-resolution TIRF-SIM microscopy videos of endocytic vesicles at the plasma membrane. In this talk, we will briefly go through our current image processing pipeline, but the main focus will be the tracking, or more precisely, the linking part. We will introduce some older and also some modern approaches and show how this task can be reformulated and solved in another domain, such as using network flows or linear integer programming.
28.2. Department meeting
11.4. Gabriela Kadlecová: Zero-cost proxies for neural network performance estimation
In neural architecture search (NAS), the task is to find the best-performing architecture for a given deep learning task. The bottleneck of the search is evaluation of network performance, as it requires training of networks, which is time-consuming. The costs can be reduced by using performace predictors, which are models that map an architecture to its trained performance. Recently, zero-cost proxies have been proposed as very fast predictors - network performance is estimated using a score that is computed by passing a single minibatch of data to an untrained network. In this presentation, I will present the zero-cost proxies and their current role in NAS, as well as my ongoing work on zero-cost proxy ensembling.
18.4. Department meeting
25.4. Věra Červíčková: Adversarial examples in Image Classification
State-of-the-art deep learning models have achieved remarkable success in pattern recognition, often outperforming humans. However, their vulnerability to adversarial attacks, where imperceptible yet intentionally harmful noise significantly reduces model accuracy, remains a significant challenge. In this talk, we will explore various methods for generating adversarial images and present techniques to defend against these malicious inputs. Focusing on a contemporary defense method, we will discuss an attack strategy that employs an evolutionary algorithm to bypass it. Our experiments reveal an interesting distinction between adversarial images generated by evolutionary methods and those created with the knowledge of gradients.
2.5. Petr Illner: Compilation Languages Based on a Relaxed Decomposability
In this talk, we propose three new complete circuit types — Bella, posBella and negBella circuits. These circuit types generalise decomposable negation normal form (DNNF) circuits in such a way that they allow a restricted form of sharing variables among the inputs of a conjunction node. Although Bella, posBella and negBella circuits have the same properties as DNNF circuits regarding the queries presented in the knowledge compilation map, we demonstrate that they differ in the cardinality queries. We also prove the unconditional strict succinctness relations between NNF, Bella, posBella, negBella and DNNF circuits.
Adam Dingle: Automatically checking proofs in controlled natural language using higher-order logic
In most interactive theorem-proving systems, the user must build a proof script that looks something like a computer program. By contrast, I have been developing a new theorem prover in which the input format looks like mathematical English, written with a limited vocabulary and grammar. The prover translates the input into formulas of higher-order logic, which it then either attempts to prove directly or can export to external higher-order provers. Starting with the Peano Postulates as axioms, the system is already capable of verifying various arithmetic identities. In this talk I will demonstrate this system and will discuss its logical foundation, its controlled natural language input format, and how the system translates and proves theorems.

Winter semester 2022/23

4.10. Cancelled
18.10. Marta Vomlelová: Active learning
Gathering labeled data is often difficult. Active learning assumes the 'easy to get' unlabeled data, for example web pages. Then, a selection strategy iteratively selects an unlabeled example to classify by an expert (oracle). Several strategies are studied. They are based on the probability distribution estimate, the risk difference estimate, the expected error reduction and others.
25.10. Jiří Švancara: Multi-agent pathfinding
The classical multi-agent pathfinding is the task of navigating a set of agents in a shared environment from their start locations to their desired goal locations. In this talk, we present the history of the problem, the most popular solving approaches, and some of our results. Mainly, we will focus on solving the classical problem optimally using reduction to SAT, then we present some practical extensions of MAPF so that it is more directly applicable to real-world problems such as autonomous intersections, warehousing, robot navigation, and much more.
1.11. Cancelled
15.11. Jakub Bulín: Few subpowers and short definitions
A relation S is pp-definable from a set of relations {R_1, R_2, ...} if it is definable by an existentially quantified conjunction of predicates, or equivalently, if it is invariant under all functions that preserve all the relations R_i. Motivated by the CSP, Berman, Idziak, Markovic, McKenzie, Valeriote, and Willard [Trans. AMS & SICOMP 2010] discovered the importance of the following property: {R_1, R_2, ...} has few subpowers, if the number of n-ary pp-definable relations is bounded by 2^p(n),for some polynomial p. Few subpowers can be seen as a generalization of "linear algebra", i.e., properties of affine subspaces over a finite field, and indeed, CSP for few subpowers can be solved by a "Gaussian eliminiation"-like algorithm. Let us consider the following related property: let us say that {R_1, R_2, ...} has short definitions if every n-ary pp-definable S can be defined by formula of length bounded by p(n), for some polynomial p. We will discuss the open question whether few subpowers imply short definitions, partial results towards a positive answer, and several related open problems.
22.11. Cancelled
29.11. Ondřej Čepek: Knowledge representation languages and knowledge compilation
In this talk we survey several types of knowledge representation languages (Boolean formulas, binary decision diagrams, list-based representations, Boolean circuits, and negational normal forms) and compilation among these languages. The talk is based mainly on the Knowledge Compilation Map framework [Darwiche and Marquis, 2002] and several recent extensions of this framework.
13.12. Cancelled
20.12. Cancelled
3.1. Cancelled