Explainable AI: Interactive User Guided Neural Network Pruning

Theoretical (Analytical):

Practical (Implementation):

Literature Work:


Overview

Neural Networks and Deep Learning have become state-of-the-art methods for various tasks. However, due to the trial and error process of building Neural Networks, it is difficult to train and use the deeper and wider architectures for prediction. Various theories exist that smaller topologies can achieve the same results as these large ones. To get to this smaller state, the Neural Network has to be pruned. Many algorithms are doing some parts automatically. However, to gain more insight into some states and prune even further, the user has to be involved. Smaller networks can then be used more efficiently to explain the workings of them.

Tasks

  • Get familiar with current Neural Network pruning techniques
  • Get familiar with interactive user guided systems
  • Take a large pre-trained model and prune it with automatic algorithms
  • Build an interactive system to further prune the network

Requirements

  • Good programming skills in Python, Javascript
    (preferable also with Pytorch or Keras / Tensorflow, D3 or WebGL)
  • Good knowledge about neural networks 
    (preferable with CNNs)

Scope/Duration/Start

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

Contact

References

  1. Frankle, J., & Carbin, M. (2018). The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks, 1–42. doi.org/arXiv:1803.03635v1