LCA: Loss Change Allocation for Neural Network Training

TL;DR

What actually happens when a model "trains"? What is each parameter's contribution to the loss change? LCA helps you catch "good" and "bad" weights during training.

Venue
In Thirty-third Conference on Neural Information Processing Systems (NeurIPS 2019).
BibTeX
@inproceedings{lan2019lca,
  title={LCA: Loss change allocation for neural network training},
  author={Lan, Janice and Liu, Rosanne and Zhou, Hattie and Yosinski, Jason},
  booktitle={Advances in Neural Information Processing Systems},
  pages={3614–3624},
  year={2019}
}
Date