Skip to content
This repository has been archived by the owner on Aug 30, 2024. It is now read-only.

Latest commit

 

History

History
46 lines (38 loc) · 4.02 KB

project.md

File metadata and controls

46 lines (38 loc) · 4.02 KB

Project

Basics

  • The topic of your project needs to be chosen until December 6th (you will need to get your topic accepted by me).
  • A short presentation (10-15min) detailing your findings is required. You will present these during the last 2 or 3 classes (depending on the number of projects). This will count for part of the grade.
  • The project should be shared as a Google Colaboratory notebook or a GitHub repository.
  • The project can be done in groups of up to three.

Proposed topics

If you're involved in a project which utilizes DL, you can build on that. The project can also be used for another course. It can also consist of (partially) replicating a research paper. The list below contains example project proposals.

Vision

  • Comparison of GAN, VAE, and WAE generative models on the SVHN dataset.
  • An interpretability method for neural networks (e.g. saliency maps, Grad-CAM, or TCAV).
  • Presentation of the problem of adversarial examples via implementation of one adversarial attack and one adversarial defense.
  • Creation of a specialized dataset (e.g. different types of birds). Fine-tuninig of a VGG variant on this dataset.
  • An algorithm for detection and valuation of Magic: The Gathering cards. Bonus: phone app.

Text

  • A generative model of the Polish language utilizing flair embeddings.
  • Comparison of simple generative models trained on different Polish writers (e.g. Mickiewicz vs. Słowacki). You can use the data from wolnelektury.pl.
  • Adapting a pretrained model for generating text for a particular domain.

Sound:

  • Shazam-like on a limited musical library.
  • Implementation of a simple generative model basing on (part of) the MAESTRO dataset.
  • An embedding for songs, thanks to which it will be easy to search for similar songs.

Games:

  • Implementation of a self-play algorithm for a game like Gomoku.
  • Replicating the DQN model and testing it on several ATARI games.
  • Utilizing a policy gradient method to solve a MuJoCo environment.
  • Creation of an environment for reinforcement learning (e.g. Candy Crush, 2048, Minesweeper) consistent with the OpenAI Gym API. Testing of several popular RL algorithms (e.g. from Stable Baselines).
    • Creation of a robotic environment with MuJoCo.

Research:

Other:

  • A generative model for Magic: The Gathering cards basing on MTGJSON (cf. RoboRosewater).
  • Creation of a dataset for a particular domain: scraping data, cleaning data, identifying potential problems, training baseline architectures, comparison with other datasets from the same domain.
  • A web service utilizing CycleGAN for automatic image modification (e.g. making people in photos look happier).