2020, Allen Goodman and Sujin Lee

• gene silencing, targeted mutations
• CCC1 recombinant protein
• α

Ideas:

• improve, add annotations—performance is dependent on localization—the better an object’s localization, the better the classifier
• extract less predictive, but far more interpretable classical features (e.g. color, shape, and texture descriptors)—especially useful for complex inference
• restructure the problem to use instance masks rather than static, 8 × 8 bounding boxes generated from regressed centroids