There is provided a system and method of automatic optimization of an examination recipe. The method includes obtaining one or more inspection images each representative of at least a portion of the semiconductor specimen, the one or more inspection images being indicative of respective defect candidates selected from a defect map using a first classifier included in the examination recipe; obtaining label data respectively associated with the one or more inspection images and informative of types of the respective defect candidates; extracting inspection features characterizing the one or more inspection images; retraining the first classifier using the first features and the label data, giving rise to a second classifier; and optimizing the examination recipe by replacing the first classifier with the second classifier; wherein the optimized examination recipe is usable for examining a subsequent semiconductor specimen.