Seminar

Our seminar

Laboratory seminars (PV273 in the course catalog) on Wednesday, 15:00-16:00, A505, FI MU, Botanická 68

The format of standard lectures: 30-40 minutes presentation + 15 minutes for questions, slides in English, presentation in English or Czech based on audience.

  • 14.9.2022
    Ing. Václav Oujezský, Ph.D.
    Connected intelligence – 6G networks
    Abstract: A new definition of the 6G network and its fundamental pillars will be presented as part of participation in the COST Action Intelligence-Enabling Radio Communications for Seamless Inclusive Interactions working group. A new communication paradigm, integrated sensing, communication-enabled intelligence, DNN model as a communication object, federated learning model for 6G, and other open calls and proposals will be discussed.
    PDF
  • 21.9.2022
    Mgr. Martin Piatka
    NAT traversal
    Abstract: In today’s Internet, NATs are ubiquitous, and it has become almost necessary for applications that employ end-to-end communication to deal with them in some way. In this talk, we will look at NATs from the perspective of a low-latency audio/video live-streaming software UltraGrid and look at practical methods for establishing connections with, or even without, the cooperation of the NAT device.
  • 5.10.2022
    doc. Mgr. Bc. Vít Nováček, PhD
    Reducing biomedical information overload by text mining and knowledge graph embeddings
    Abstract: Information overload is arguably relevant to virtually any human activity these days. However, it is particularly pertinent to life sciences and healthcare. This motivates the specific instances of the information overload problem that will be addressed in the talk: 1) The vast breadth and depth of published life science articles that can hardly be utilized in a focused and exhaustive manner. 2) The largely untapped potential of networked biomedical data for making semi-automated discoveries. I will review one approach to tackling the first challenge that is based on ontology learning from text. The other, more recent solutions I will address the second challenge by means of relational machine learning, applied to link prediction in networked biomedical data. The talk will then conclude with an overview of two ongoing research collaborations that explore clinical applications of the presented techniques.
  • 12.10.2022
    Eric R. Anschuetz
    Interpretable quantum advantage in neural sequence learning
    Abstract: Neural networks implemented on quantum computers have been widely studied in recent years, given their potential practical utility and the computational power of quantum computers. However, no unconditional analytic results on the expressive power of quantum neural networks have been shown. Here, we study the relative expressive power between a broad class of neural network sequence models and a class of quantum recurrent models. We explicitly show which physical phenomena yield a memory separation in the expressivity of the two model classes. We then use this intuition to numerically study the relative performance of our introduced model on a standard translation data set. In doing so, we demonstrate that our introduced quantum models are able to outperform state of the art traditional machine learning models even in practice.

    Link for the associated paper: https://arxiv.org/abs/2209.14353

  • 19.10.2022
    Canceled
  • 26.10.2022
    About Sitola laboratory
  • 2.11.2022
    Canceled: MetaCentrum seminar
  • 9.11.2022
    Mgr. Filip Petrovič
    Generating and storing large autotuning configuration spaces
    Abstract: When optimizing the performance of GPU compute kernels, a programmer has to make many decisions, such as how to arrange data structures in memory, block data access to optimize caching, or which factor to use for loop unrolling. All of these decisions affect kernel performance. Autotuning is a technique that automatically explores combinations of such decisions to find the best performing version (also called a configuration). All the individual configurations make up a configuration space. In this lecture, we will look at how to generate and store large configuration spaces. That is, spaces of autotuning problems with a large number of optimization decisions and/or constraints.
  • 16.11.2022
    Bc. Adam Hájek
    Experience with large-scale NLP-inspired neural networks applied in another domain
    Abstract: Compound identification in mass spectrometry, i.e., finding a formula (e.g., „CCCCCc1cc(O)c2c(c1)OC(C)(C)[C@@H]1CCC(C)=C[C@@H]21“) matching an experimentally measured mass spectrum (list of few dozens of pairs of numbers — mass-intensity) is addressed by similarity search in spectra databases in conventional methods. However, these databases do not cover all compounds of interest (approx. 250,000 known spectra vs. 1,000,000,000 compounds known to exist and up to 10^60 theoretically possible compounds). Therefore „de novo“ methods recently gained interest.
    These approaches try to state the problem as „translation“ from the „language of spectra“ to the „language of formulae“, and to apply methods known to work with natural languages. It appears that the reverse direction (formula -> spectrum) is fairly easy. With the existing 250k database, a straightforward NN can be trained to achieve reasonable accuracy. On the contrary, the spectrum->formula direction is far more difficult, larger and more complex NN are required, and a larger training set is needed in turn. This can be generated by applying the reverse model on a huge (millions) set of feasible formulae. Although such generated spectra are only moderately accurate, they are sufficient to train the large model.
    We will give a more detailed description of the method, demonstrate its functionality, and share our experience with its independent implementation and adaptation to the specific needs of MU Recetox laboratory.
  • 23.11.2022
    RNDr. David Střelák
    Getting things done. Or not.
    Abstract: With cutting-edge technology and tools, we should be able to produce more work in less time.
    However, we often find ourselves overwhelmed by tasks, requests, pending emails, and life.
    Is there a way out?
    David Allen proposes a personal productivity system called Getting Things Done (GTD). The main idea of the system is to get things out of your mind so you can focus on one thing at a time.
    In this presentation, we will go through his system and see how I applied it in my life.
    Since the GTD system also considers the known time limitation, the presentation will also include thoughts from Four Thousand Weeks: Time Management for Mortals by Oliver Burkeman.
  • 30.11.2022
    Ing. Tomáš Martínek, Ph.D.
    In-band Network Telemetry – configurable data plane monitoring framework
    Abstract: With the advent of programmable network switches, the P4 language and SDN network architecture, new technologies for network traffic monitoring are emerging. One such technology is In-band Network Telemetry (INT), which enables precise monitoring of individual network flows directly at the data plane level. The presentation will introduce the basic principles and models on which INT technology is based. The presentation will be complemented by practical experience gained during the construction of the European measurement testbed within the GEANT 4 project.
  • 7.12.2022
    Mgr. Martin Friák, Ph.D.
    Theoretical approaches to acoustic emissions in materials
    Abstract: The analysis of acoustic emissions represents a very important tool in modern materials science. Impending catastrophic failure of mechanically highly stressed components may be detected well in advance due to the fact that the formation and motion of internal defects result in acoustic emissions. Unlike numerous surface-focused characterization techniques, acoustic emissions provide an insight into the depth of the studied material. Measurement of acoustic emissions plays therefore a vital role among safety measures in critical infrastructures such as nuclear power plants. However, the analysis of acoustic emissions is challenging for several reasons. First, at the macroscale, the outcome of minutes-long testing is represented by dozens of gigabytes of data, which is difficult for a human operator to handle. Second, at the nano-scale, fracture-related atomistic processes are nearly impossible to measure and very difficult to compute. While information science offers tools to solve the first issue (see the master thesis of Mgr. Marek Pastierik co-supervised by Dr. Tom Rebok, Faculty of Informatics, Masaryk University, Brno, 2022), the second problem is still not solved. My seminar lecture will therefore outline selected quantum-mechanical and atomistic modeling approaches to (i) vibrations of atoms in general and (ii) acoustic emissions in particular. As a modern trend in computational materials science, inter-atomic potentials obtained by machine learning from molecular-dynamics simulations will be discussed as well.
  • Tuesday 13.12.2022 16:00
    Sitola Christmas seminar

Past seminars

Contact: Hana Rudová

(c) 2021 SITOLA, Administration