Computer Vision News - August 2021

2 Summary AI Rese ch 8 5) Identify and evaluate hand-engineered features. These are created by computing mean RGB intensities within 10 concentric circles (5mm apart) around the fovea. These are engineered with the purpose of evaluating previous hypotheses by the image translation method of the essentiality of the fovea area for the model predictions. When used in combination with the presence of hard exudates to train two simple classifiers (a linear SVM with l1 regularization and an MLP with 1 hidden layer), these perform similarly to the model M (CNN on raw pixels single task on cropped image) for the task of DME classification . To summarize, the most important contributions of this work include 1) the creation of a framework to convert predictions from explanation techniques to a mechanism for discovery; 2) a proof that generative models in combination with black-box predictors can generate hypotheses without human priors that can be critically explained; and 3) a case study on classification for retinal images predicting DME. Some nice extra images from this work below… Examples of converting CFP images from DME to no-DME predictions (model f).