PhD Research Seminar: Educational games, sentiment analysis, molecular-dynamics simulation
When: April 6, 18:10–21:00
First talk: Educational Game Analysis Using Intention and Process Mining
Speaker: Konstantin Nikitin, second-year PhD student, Faculty of Computer Science
Gamification is a common practice that however leads to more complex and difficult processes to model. Goal-oriented process mining and intention mining in particular seem to be efficient for such task. However, while existing methods can help reveal intentions within an educational game, they are too abstract in the context of behaviour analysis. Activity-oriented process modelling techniques are precise, but tend to discover so-called "spaghetti-like" models. So we propose a formalism that combines both perspectives to model unstructured processes in a hierarchical way. Initialization of model properties and some assumptions about the resulting two-perspective formalism features will be presented. We will also discuss possible ways of conformance checking for the described model.
Second talk: Language Models Fine-Tuning for Sentiment Analysis of Russian-language Content
Speaker: Sergei Smetanin , second-year PhD student, School of Business Informatics
Nowadays, fine-tuning of pre-trained language models achieves competitive results in a range of natural language processing tasks, including sentiment analysis. In this talk, we identify fine-tuning baselines for sentiment analysis of Russian-language textual content. We fine-tunned Multilingual Universal Sentence Encoder and Multilingual Bidirectional Encoder Representations from Transformers on five publicly available sentiment datasets in Russian and obtained strong or even new state-of-the-art results. We made trained models and code samples publicly available for the research community.
Third talk: Parallel GPU Algorithm for Molecular-Dynamics Simulation with Four-Point Water Model
Speaker: Vsevolod Nikolskiy, second-year PhD student, HSE Tikhonov Moscow Institute of Electronics and Mathematics
Classic molecular dynamics is a very powerful modeling method that plays a significant role in many areas of modern science. Applications that implement this method produce a huge part of the computational load on supercomputers around the world. For this reason, the development of new algorithms for using current and promising computing technologies to increase performance is very much in demand. Water molecules are used in a large number of models. Three-point water models are simple and intuitive, but they have a number of disadvantages when used in classical molecular dynamics. Adding a massless virtual charge to each water molecule can significantly improve the distribution of electrostatic fields and makes it possible to model water in a wide range of parameters. Such a four-point model TIP4P is more difficult to program and compute. It is required to update the coordinates of the virtual charges at each timestep, calculate the interactions, and distribute the contribution from virtual particles between the real atoms. The adaptation of the implemented TIP4P CPU code for calculations on graphics accelerators is associated with some algorithmic difficulties and could not be done in the LAMMPS package yet. In this work, a new algorithm was created for calculating TIP4P as part of the universal package of classical molecular dynamics LAMMPS. The code was implemented in the GPU module as a pair style lj/cut/tip4p/long/gpu. It completely inherits the interface of the standard lj/cut/tip4p/long and it is compatible with the pppm/tip4p. The code allows the use of Nvidia and AMD GPUs. The new code reduces the time-to-solution by about 60-70% on typical systems and makes it possible to solve new problems and obtain the original physical results.