ABSTRACT
Gradients of a deep neural network’s predictions with respect to the inputs are used in a variety of downstream analyses, notably in post hoc explanations with feature attribution methods. For data with input features that live on a lower-dimensional manifold, we observe that the learned function can exhibit arbitrary behaviors off the manifold, where no data exists to anchor the function during training. This leads to a random component in the gradients which manifests as noise. We introduce a simple correction for this off-manifold gradient noise for the case of categorical input features, where input values are subject to a probabilistic simplex constraint, and demonstrate its effectiveness on regulatory genomics data. We find that our correction consistently leads to a significant improvement in gradient-based attribution scores.
Competing Interest Statement
The authors have declared no competing interest.
Footnotes
↵* majdand{at}cshl.edu and koo{at}cshl.edu