Explaining an image classifier with a generative model conditioned by uncertainty

TL;DR

Preliminary experiments on synthetic data and a corrupted version of MNIST dataset illustrate the idea to condition a generative model by a given image classifier uncertainty in order to analyze and explain its behavior.

Abstract

We propose to condition a generative model by a given image classifier uncertainty in order to analyze and explain its behavior. Preliminary experiments on synthetic data and a corrupted version of MNIST dataset illustrate the idea.

Figures (4)

• Figure 1: Training process and structure of the generator.
• Figure 2: Qualitative and quantitative results for moons dataset. Uncertainty conditioning works well; the MSP condition corresponds roughly to the real MSP.
• Figure 3: Samples of images generated with MSP condition fixed at $1$ (top) and $0.7$ (bottom). Above each image is shown the classifier prediction and probability. Images at the bottom look harder, and the classifier is more uncertain.
• Figure 4: MSP condition (in) vs. MSP computed by classifying the generated data (out).