When less is more: Simplifying inputs aids neural network understanding

TL;DR

We made a knob which when turned makes the input images simpler and simpler. With that we study the network's response in different scenarios. Kinda like pruning only not on weights but on input; except we don't just remove pixels, but learn to generate simpler images.

Venue
arXiv
BibTeX
@article{schirrmeister2022more,
  title={When less is more: Simplifying inputs aids neural network understanding},
  author={Robin Tibor Schirrmeister and Rosanne Liu and Sara Hooker and Tonio Ball},
  year={2022},
  eprint={2201.05610},
  archivePrefix={arXiv},
  primaryClass={cs.LG}
}
Date