Training and Generating Neural Networks in Compressed Weight Space
Abstract
The inputs and/or outputs of some neural nets are weight matrices of other neural nets. Indirect encodings or end-to-end compression of weight matrices could help to scale such approaches. Our goal is to open a discussion on this topic, starting with recurrent neural networks for character-level language modelling whose weight matrices are encoded by the discrete cosine transform. Our fast weight version thereof uses a recurrent neural network to parameterise the compressed weights. We present experimental results on the enwik8 dataset.
Cite
Text
Irie and Schmidhuber. "Training and Generating Neural Networks in Compressed Weight Space." ICLR 2021 Workshops: Neural_Compression, 2021.Markdown
[Irie and Schmidhuber. "Training and Generating Neural Networks in Compressed Weight Space." ICLR 2021 Workshops: Neural_Compression, 2021.](https://mlanthology.org/iclrw/2021/irie2021iclrw-training/)BibTeX
@inproceedings{irie2021iclrw-training,
title = {{Training and Generating Neural Networks in Compressed Weight Space}},
author = {Irie, Kazuki and Schmidhuber, Jürgen},
booktitle = {ICLR 2021 Workshops: Neural_Compression},
year = {2021},
url = {https://mlanthology.org/iclrw/2021/irie2021iclrw-training/}
}