Запрошуємо переглянути захисти дипломних магістерських робіт студентів програми «Науки про дані», які пройдуть 5 червня 2020 року. Розклад та теми робіт:

• 14:00, Володимир Лут, «Пошук архітектур нейронних мереж: імовірнісний підхід»
  • In this project, we introduce the Bayesian Optimization (BO) implementation of the NAS algorithm that is exploiting patterns found in most optimal unique architectures sampled from the most popular NAS dataset and benchmarking tool NASbench-101. The proposed solution leverages a novel approach to path-encoding and is designed to perform reproducible search even on a relatively small initial batch obtained from the random search. This implementation does not require any special hardware, it is publicly available.
• 14:40, Антон Щербина, «Покращенння контрольованості генерації тексту»
  • Many models could generate text conditioned on some context, but those approaches don’t provide us with the ability to control various aspects of the generated text (e.g., sentiment). To address this problem, Variational Autoencoder is typically used because they give the ability to manipulate in latent space and, in this way, control text generation. However, it has been shown that VAE with strong autoregressive decoders, which are used for text modeling, faces posterior collapse problem. We think that one of the reasons why this problem occurs is a restrictive gaussian assumption we made about approximate posterior. In this work, we want to apply well-known approaches based on Normalizing Flows to improve approximate posterior for text modeling and check if it can help avoid posterior collapse.
• 15:20, Костянтин Лапчевський, «Передбачення властивостей кристалічних структур»
  • Crystalline structures are vital to the modern technology. Yet, we are still only starting to figure out how to properly estimate their directional properties using machine learning techniques. In order to improve that, I build upon the theory and codebase of Euclidean Neural Networks (networks equivariant to 3D rotations). The main contributions of this work are: a derivation of the decompositon/reconstruction equations of elastic tensor that enables using it as a train target, optimized CUDA implementation of the core operation PeriodicConvolution that makes it fast and scalable, and an analysis of the trends of geometric structures and electronic properties of the crystal in Materials Project Database and how these trends impact hyper-parameters for convolutional neural network architectures such as Euclidean Neural Networks.

Захисти будуть відбуватися онлайн. Долучитися до перегляду можна через сервіс Zoom.