Validation and generalization of pixel-wise relevance in CNNs trained for face classification

The increased use of convolutional neural networks for face recognition in science,
governance, and broader society has create an acute need for methods that can
codify how these ’black box’ decisions are made. To be interpretable and useful
to humans, such a method should convey a model’s learned classification strategy
in a way that is robust to random initializations or spurious correlations in input
data. In this work, we applied the decompositional pixel-wise attribution method
of layer-wise relevance propagation (LRP) to this end, quantifying pixel-wise
information use of several classes of VGG-16 models trained for face recognition.
We evaluated the models on relevance-based occluded images to analyze how
each model’s relevance distributions vary with and generalize across the type of
pretraining dataset (ImageNet or VGGFace), the finetuning task (gender or identity
classification), and random initializations. We found that relevance distributions
produced by these various types of VGG-16 models prove generally stable across
random initializations and can generalize across finetuning tasks. However, there
is much less generalization across types of pretraining data. So, we find that
LRP relevance distributions generalize asymmetrically and to varying degrees
across these parameters, suggesting it may be possible to find an underlying set of
important facial image pixels that are relevant to most CNNs and tasks.