This challenge is inspired by research in human computation like "beat the machine" by Attenberg, Ipeirotis and Provost (2011) and programs like vulnerability rewards.

Participants in CATS4ML will discover and submit adverse images, which we call AI blindspots. Participants will invent new and creative ways to explore an existing publicly available benchmark dataset and discover blindspot examples guided by a list of preselected target labels.

AI Blindspots are not the typical adversarial examples defined in ML literature by Goodfellow (2015) as machine manipulated images aimed to fool machines and to be imperceptible to humans,

AI blindspots are unmanipulated (real) images for which humans can reliably agree on a label but most AI models would disagree.

AI blindspots are unknown unknowns, e.g. images with visual patterns that can not be easily distinguished by AI models because they are:

• rare, e.g. images of ‘dog breeds’ that may be rare on the Internet and AI models might be confident but wrong in not labeling these images with ‘Catalburun’ (a rare dog breed);
• tricky, e.g. images that have a tricky combination of ‘objects’ and ‘positions’ like cats that looks like a dog, but unambiguous from the context, where AI models might be confident but wrong in labeling an image with ‘cat’ instead of ‘dog’;
• both rare and tricky. e.g. images of race cars with heat-induced visual distortion from underlying concrete; this are common, archetypal images in Hollywood genres as cars approach from a distant horizon, but the visual mirage could lead to a very odd shaped car and confuse AI models trained to predict `car` from more traditional camera perspectives.

In the CATS4ML challenge, finding AI blindspots is about using human intuition to find the unknown unknowns, unlike traditional active learning techniques, which utilize signals such as machine confidence to explore what the model “thinks” it does not know (i.e. the known unknowns).

Check Participate to learn how to start contributing.