Neural Network Attributions: A Causal Perspective

Chattopadhyay, Aditya and Manupriya, Piyushi and Balasubramanian, Vineeth N and et al, . (2019) Neural Network Attributions: A Causal Perspective. In: Proceedings of the 36th International Conference on Machine Learning (ICML), 9-15 June 2019, California, U S.

Full text not available from this repository. (Request a copy)

Abstract

We propose a new attribution method for neural networks developed using firstprinciples of causality (to the best of our knowledge, the first such). Theneural network architecture is viewed as a Structural Causal Model, and amethodology to compute the causal effect of each feature on the output ispresented. With reasonable assumptions on the causal structure of the inputdata, we propose algorithms to efficiently compute the causal effects, as wellas scale the approach to data with large dimensionality. We also show how thismethod can be used for recurrent neural networks. We report experimentalresults on both simulated and real datasets showcasing the promise andusefulness of the proposed algorithm.

[error in script]

IITH Creators: