A Causal View on Robustness of Neural Networks

TL;DR

This paper introduces a causal perspective on neural network robustness, proposing a deep causal manipulation augmented model (deep CAMA) that explicitly models cause manipulations, leading to improved robustness and disentangled representations.

Contribution

The paper presents deep CAMA, a novel causal modeling approach that enhances neural network robustness and achieves disentangled representations, extending robustness analysis beyond traditional classification.

Findings

Deep CAMA outperforms standard neural networks against unseen manipulations.

The model achieves disentangled representations separating manipulations from other causes.

Data augmentation and fine-tuning further improve robustness.

Abstract

We present a causal view on the robustness of neural networks against input manipulations, which applies not only to traditional classification tasks but also to general measurement data. Based on this view, we design a deep causal manipulation augmented model (deep CAMA) which explicitly models possible manipulations on certain causes leading to changes in the observed effect. We further develop data augmentation and test-time fine-tuning methods to improve deep CAMA's robustness. When compared with discriminative deep neural networks, our proposed model shows superior robustness against unseen manipulations. As a by-product, our model achieves disentangled representation which separates the representation of manipulations from those of other latent causes.