English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

The increased use of convolutional neural networks for face recognition inscience, governance, and broader society has created an acute need for methodsthat can show how these 'black box' decisions are made. To be interpretable anduseful to humans, such a method should convey a model's learned classificationstrategy in a way that is robust to random initializations or spuriouscorrelations in input data. To this end, we applied the decompositionalpixel-wise attribution method of layer-wise relevance propagation (LRP) toresolve the decisions of several classes of VGG-16 models trained for facerecognition. We then quantified how these relevance measures vary with andgeneralize across key model parameters, such as the pretraining dataset(ImageNet or VGGFace), the finetuning task (gender or identity classification),and random initializations of model weights. Using relevance-based imagemasking, we find that relevance maps for face classification prove generallystable across random initializations, and can generalize across finetuningtasks. However, there is markedly less generalization across pretrainingdatasets, indicating that ImageNet- and VGGFace-trained models sample faceinformation differently even as they achieve comparably high classificationperformance. Fine-grained analyses of relevance maps across models revealedasymmetries in generalization that point to specific benefits of choiceparameters, and suggest that it may be possible to find an underlying set ofimportant face image pixels that drive decisions across convolutional neuralnetworks and tasks. Finally, we evaluated model decision weighting againsthuman measures of similarity, providing a novel framework for interpreting facerecognition decisions across human and machine.

Validation and generalization of pixel-wise relevance in convolutional neural networks trained for face classification