Schedule

Every month a speaker from one of the organizing research groups will present their work.

Abstracts

13 June 2022

Speaker: Masoumeh Mansouri

Combining Task and Motion Planning: Challenges and Guidelines

Combined Task and Motion Planning (TAMP) is an area where no one-fits-all solution can exist. Many aspects of the domain, as well as operational requirements, have an effect on how algorithms and representations are designed. Frequently, trade-offs have to be made to build a system that is effective. In this talk, I present five research questions that I believe need to be answered to solve real-world problems that involve combined TAMP. Then, I will show which decisions and trade-offs we made with respect to these research questions, and illustrate these based on our work on several application domains including drill planning for open-pit mines, multi-hauler planning for quarrying and assembly planning for industrial manipulators.

16 May 2022

Speaker: Adam Pease

Ontology for Machine Learning

Symbolic and statistical methods have often been considered distinct approaches to AI. Some significant recent work at CVUT by (Urban et al.) has been in applying machine learning to improve theorem proving, especially in the area of mathematical proofs. We might call this machine learning applied to symbolic AI methods. There has also been some work that is often called neuro-symbolic AI using symbolic representations to aid machine learning. In this talk, we discuss the Suggested Upper Merged Ontology (SUMO) and its potential to support machine learning AI methods for a variety of applications.

25 April 2022

Speaker: Jelena Mitrović

Computational Rhetoric in Social Media and Law

Computational Rhetoric, i.e. processing of non-literal natural language and detection of rhetorical figures, applied to Sentiment Analysis, Offensive language modelling and detection, Legal Tech, Argumentation Mining and other areas of NLP, has the potential to facilitate deeper, automatic understanding in natural language. In this talk, we show how computational modelling of rhetorical figures and the relations that exist between their building blocks, can help us come closer to achieving that.

28 March 2022

Speaker: Erik Derner

Data-efficient methods for model learning and control in robotics

In this talk, I will present efficient methods for data-driven model learning in robotics. Models facilitate simulations, analysis of the robot's behavior, decision making, and design of automatic control algorithms. However, obtaining informative data for constructing dynamic models can be difficult, especially when the models are to be learned during task execution. To this end, symbolic regression proves to be a suitable method to automatically build such models. This technique, based on genetic programming, is able to construct accurate and compact models in the form of analytic equations. It represents a viable alternative to popular data-hungry methods such as deep neural networks, which typically produce black-box models that are much larger in the number of parameters. One of the challenges in continual model learning is posed by the large amount of data collected by the robot. I will present a comparison of sample selection methods and show that symbolic regression can be used to construct accurate models from very small informative training sets. Another challenge consists in that data-driven model learning techniques often produce models that are accurate, but do not comply with the physical properties of the robot. I will explain how symbolic regression can be naturally extended to take into account the physical constraints and include the prior knowledge in the form of a theoretical or empirical model into the model construction process.

21 February 2022

Speaker: Petr Marek

Conversational AI Alquist: Teaching Computers to Surpass Human Baseline in Post-Pandemic Pub Debates

How can we develop a conversational AI that conducts coherent and engaging conversations? What are the state-of-the-art approaches, and are they applicable in practice? And where is the research of conversational AI heading next? Petr Marek, a member of the Alexa Prize-winning team Alquist, will answer those questions. Moreover, he will show you the limitations of recent large scale language models like OpenAI's GPT-3, Google's Lamda or Facebook's Blenderbot and how the team Alquist addresses them in their experiments and practice. Lastly, we will answer the question of when will some kind of digital persona invite us to talk over a non-digital beer.

24 January 2022

Speaker: Přemysl Šůcha

Minimizing the weighted number of tardy jobs on a single machine: Strongly correlated instances

This work addresses a single machine scheduling problem minimizing the weighted number of tardy jobs, where each job is characterized by processing time, due date, deadline, and weight. It is known from the existing literature that so-called strongly correlated instances, i.e., instances where each job has the weight equal to its processing time plus a constant, are significantly harder to solve compared to instances without this relation. In this work, we extend an exact algorithm proposed in Baptiste et al. (2010) with the aim of solving the strongly correlated instances significantly faster. The main improvement is the new integer linear programming model for strongly correlated instances utilizing a decomposition according to the number of tardy jobs. Other proposed improvements are tighter lower and upper bounds which can be applied to all types of instances. The best-known algorithm proposed in Baptiste et al. (2010) cannot solve all instances with 250 jobs to the optimum within an hour. On the same hardware, our relatively simple improvements implemented into the algorithm proposed by Baptiste et al. enable solving all examined strongly correlated instances to the optimum within an hour for up to 5,000 jobs and reduce the computational time on other instances as well.

8 November 2021

Speaker: Jan Zahálka

Hybrid Intelligence: Interactive and Semi-Autonomous Learning

Hybrid intelligence (HI) is an emerging discipline that ultimately aims to combine human intellect with AI to unlock solving tasks that would be difficult or impossible to solve by a human or a machine alone. In more immediate terms, HI strives to intelligently assist humans with complex tasks that are difficult to solve by fully automatic and/or autonomous AI. My research within this domain is in interactive and semi-autonomous learning: ensuring the machine learns from a user's interactions and adapts its model to suit the user's mental model of the task at hand. This talk is an introduction to HI, my existing work, and future research.

6 April 2021

Speaker: Torsten Sattler

Visual Localization: Algorithms, Applications, and Privacy

Visual localization is the problem of determining the exact position and orientation from which a given image was taken in a known scene. Solutions to the visual localization problem are core components in applications such as self-driving cars, autonomous robots, and augmented / mixed reality systems. This talk consists of three parts covering recent work to be presented at CVPR 2021: we start with a recent interesting application of visual localization, namely estimating the body pose of a human in a scene using only ego-centric data. To this end, IMU sensors mounted on the human body and a camera mounted on the head are used to jointly reason about the position of the human in the scene and their body pose. Next, we discuss one of the main components of any localization system: establishing correspondences between pixels in the given image and the scene representation used by the localization algorithm. Finally, we cover privacy aspects of visual localization algorithms: many state-of-the-art approaches rely on an explicit 3D model of the scene, often in the form of sparse point clouds. However, recent work has shown that detailed images can be recovered from such point clouds. Lifting 3D points to 3D lines prevents existing image recovery approaches while still facilitating camera pose estimation. While prior work has claimed that such a lifting approach is privacy-preserving, we show that this is not necessarily the case and that 3D points can be (accurately) recovered from line clouds purely based on geometric reasoning.

8 May 2021

Speaker: Martin Suda

Boosting Automated Reasoning using Machine Learning

Clause selection is one of the key decision points within the saturation-style architecture of automated theorem provers (ATPs) for first-order logic. I will describe how machine learning (ML) can be used to greatly improve the clause selection heuristics and thus the prover performance. I will then put this prototypical example of the ATP+ML synergy into a broader context.

12 April 2021

Speaker: Jonáš Kulhánek

Transformers: From Language to Images

Deep neural networks based only on attention, called transformers, achieved superior performance in most natural-language processing domains, replacing previously used recurrent neural networks. Recently, they showed promising results on computer vision domains such as image recognition and object detection. In this talk, I will introduce their architecture and talk about two applications that we currently work on: end-to-end task-oriented dialogue modelling and neural 3D-scene representation.

8 March 2021

Speaker: Barbora Hudcová

Complex Systems As a Possible Path to AI

There are many classes of discrete dynamical systems whose visualizations seem very intriguing, with structures spontaneously growing and further growing in complexity. These systems with such fascinating complex dynamics are hypothesized to be potentially good models for simulating some form of artificial evolution, possibly leading to AGI. In this talk, we discuss a famous problem of how to formally classify the dynamics of discrete dynamical systems and separate the ones with complex dynamics from the rest. We present a novel classification based on studying the asymptotic behavior of the average computation time in a given system before entering a loop. With it, we were able to identify a critical region of behavior that corresponds to a phase transition from ordered behavior to chaos across various classes of dynamical systems

25 January 2021

Speaker: Josef Šivic

Automatic visual recognition: from pixels to machines that see, reason and act

Building machines that can automatically understand complex visual inputs is one of the central problems in artificial intelligence with applications in autonomous robotics, automatic manufacturing or healthcare. The problem is difficult due to the large variability of the visual world. In this talk, I will present some of our contributions to the recent progress in automatic visual recognition and discuss some of the key open challenges. The talk should be accessible to non-specialists in computer vision and will try to make connections to other research areas and groups at CIIRC including robotics, language understanding, and reasoning.