PhD Research Seminar: Sampling Techniques in Rendering, Pretrained Models for Source Code Processing, Agent Navigation and Coordination
When: June 15, 9:30–10:50
First talk: Improving Sampling Techniques in Rendering
Speaker: Sofya Ivolgina, second-year Master's student (combined Master's-PhD track), Faculty of Computer Science
Rendering is the process of producing photo-realistic images and videos by computer. Generally, there are two types of rendering systems: scenes can be prerendered or generated in real time. In our work, we focus on online rendering, which is extensively used in video games and other applications with dynamically drawing scenes.
The main time-consuming process in rendering is to evaluate the brightness of each pixel in the picture considering the direct and indirect light of the scene. Such calculation is described mathematically by the rendering equation as a high-dimensional integral over all light sources. One of the principal techniques to approximate such an integral is the Monte-Carlo method, which suffers from a high variance of the estimate.
The main purpose of this work is to compare known rendering methods and their sampling techniques in terms of the quality of inferred images and computational time.
Second talk: Pretrained Models for Source Code Processing
Speaker: Sergey Troshin, second-year Master's student (combined Master's-PhD track), Faculty of Computer Science
Deep learning models are widely used for solving challenging code processing tasks, such as code generation or code summarization. Recently general-purpose pretrained models such as CodeBERT or CodeT5 have been shown to outperform task-specific models in many applications. While pretrained models are known to learn complex patterns from data, they may fail to understand some properties of source code. To test diverse aspects of code understanding, we introduce a set of diagnosting probing tasks. We show that pretrained models of code indeed contain information about code syntactic structure and correctness, the notion of namespaces, code readability, and natural language naming, but lack understanding of code semantics. We also investigate how probing results are affected by using code-specific pretraining objectives, varying the model size, or finetuning.
Moreover, we investigate the subtokenization techniques for pretrained models of code and study the effect of subtokens granularity on the downstream performance. We propose subtokenziation that reduces average length by 17–40% without downstream performance drop and show that carefully chosen subtokenization may significantly improve the quality by 0.5–2%, possibly with some length increase.
Third talk: Individual Agent Navigation in an Uneven Environment and Coordination of Mobile Agents in Multi-agent Navigation
Speaker: Stepan Dergachev, second-year Master's student (combined Master's-PhD track), MIEM
We consider two topics from the field of navigation of mobile agents.
In the first part of the talk, the local path following problem of an individual agent (mobile robot) in an uneven environment is considered. We propose to use the Model Predictive Path Integral (MPPI) control method and suggest novel cost functions for MPPI in order to adapt it to elevation maps and motion through unevenness.
In the second part, we discuss the problem of coordination in decentralized navigation of a group of mobile agents. To solve this problem, a method of integrating centralized planning algorithms into the decentralized navigation pipeline and a new centralized planning algorithm on non-regular graphs that takes into account the size of agents are proposed.
Трошин Сергей Дмитриевич
Центр глубинного обучения и байесовских методов: Стажер-исследователь