Lakhasly

Our approach is based on 4 fundamental steps that encompass both technical and interpretation tasks: (i) Train a classifier on a given image dataset to perform the desired machine learning task.If the overall accuracy of the trained classifier is high, then the information for visual assessment is assumed to be present in those images.

Our approach is based on 4 fundamental steps that encompass both technical and interpretation tasks: (i) Train a classifier on a given image dataset to perform the desired machine learning task. If the overall accuracy of the trained classifier is high, then the information for visual assessment is assumed to be present in those images. (ii) Train a “StylEx” model—a StyleGAN-based image generator guided by the classifier—on this dataset of images. The loss function encourages the GAN to generate images which both resembles the original input and are correctly classified by the classifier, forcing the generator to put an emphasis on the visual attributes which are most relevant for classification. (iii) Automatically extract the top attributes which affect the classifier. (iv) Generate a visualization of these attributes for subject matter experts’ evaluation, discussion, and consensus. As part of this framework, an interdisciplinary panel including both data domain experts and social scientists then interpret the attributes for consistency with prior literature and reason about whether any could be novel discoveries. Importantly, the interdisciplinary nature of the panel reduced blind spots based on individual expertise.