Improving performance of deep learning models with axiomatic attribution priors and expected gradients — mentioned by gabeerion

Thu, 27 Jun 2019 13:14:42 GMT

I wanted to share this work from a machine learning paper (https://arxiv.org/abs/1906.10670) we recently submitted. TL;DR - the idea is that there has been a lot of recent research on explaining neural networks by attributing importance to each input feature. We go one step farther and incorporate attribution priors - prior beliefs about what these feature attributions should look like - into the training process. We develop a fast, differentiable new feature attribution method called expected gradients, and optimize differentiable functions of these feature attributions to improve…

Paper and discussion history. A mention is not an endorsement.

Improving performance of deep learning models with axiomatic attribution priors and expected gradients · HackerPapers discussions

Improving performance of deep learning models with axiomatic attribution priors and expected gradients — mentioned by gabeerion