Explainable Time Series Generation

Theoretical (Analytical):

Practical (Implementation):

Literature Work:


Overview

Machine learning and especially deep learning introduced a new potential to generate new data-based loosely on available data by incorporating various learning and architecture techniques. However, in most cases these are only used for images or text, but it is also possible to use them for time series. This project evolves around a generator for time series. In the first step, a Generative Adversarial Networks (GAN [1]) should be trained to generate time series. The second step incorporates certain conditions into the GAN (C-GAN [2]) to enable to steer the generation into a direction. The third and last step introduces XAI methods such as LRP [3] to highlight the latent space of the generator and the decisions of the discriminator. The final result consists of a working prototype to generate time series with specific conditions using deep learning.

Tasks

  • Get familiar with current GAN approaches
  • Train a custom GAN for time series generation
  • Extend your GAN with conditions to steer generation
  • Get familiar with current XAI methods
  • Extend your GAN with current XAI methods
  • Design and implement a visual interface for the GAN model and its explanations

Requirements

  • Good programming skills in Python, Javascript
    (preferable also with Pytorch or Keras / Tensorflow, D3 or WebGL)
  • Good knowledge about neural networks 
    (preferable but not fully necessary)

Scope/Duration/Start

  • Scope: Bachelor/Master
  • Duration: 6 Month Project, 3 Month Thesis (Bachelor) / 6 Month Thesis (Master)
  • Start: immediately

Contact

References

  1. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., ... & Bengio, Y. (2014).
    Generative adversarial nets.
  2. Mirza, M., & Osindero, S. (2014).
    Conditional generative adversarial nets.
  3. Bach S, Binder A, Montavon G, Klauschen F, Müller K-R, Samek W (2015).
    On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation.