Under the Hood of Neural Networks: Characterizing Learned Representations by Functional Neuron Populations and Network Ablations
preprint
Studying learned representations through functional neuron populations, activation patterns and network ablations.
Resources
Publication history
Preprint: 2 April 2020. Acceptance at ICAI 2021 was announced by the authors’ institution in May 2021. The citation below refers to the publicly available arXiv manuscript.
Summary
This study groups neurons by their activation behavior and uses network ablations to investigate the functional roles of learned representations. It examines how activation selectivity relates to a neuron’s contribution to task performance, showing the limitations of relying on either property alone to assess importance. The approach connects analyses of individual neurons with the behavior of functional neuron populations.
Citation
@misc{meyes2020underthehood,
author = {Meyes, Richard and Waubert de Puiseau, Constantin and Posada-Moreno, Andres and Meisen, Tobias},
title = {Under the Hood of Neural Networks: Characterizing Learned Representations by Functional Neuron Populations and Network Ablations},
year = {2020},
eprint = {2004.01254},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2004.01254}
}