Subject Media Associations

Subject-media associations are the primary mechanism by which information of various forms is associated with visual media in the Cogniac system. All model predictions and user feedback result in subject-media associations that vary in probability based on the strength of the association. As a reminder a subject media association probability of 1 corresponds with a definite positive association (subject present) and a probability of 0 corresponds with a definite negative association (subject not present).

The Cogniac system is constantly assessing the accuracy of both user feedback and application model predictions (based on user feedback). User feedback is cross-validated against other user feedback to assess individual user feedback accuracy. Application model predictions are validated against user feedback to assess model accuracy. Both of these assessments occur within a Bayesian statistical framework that treats both users and models almost identically. The only difference is that there is a baseline assumption that on average our users are better than random, e.g. not evil. We prefer not to have to relax that assumption...

Based on these assessments, after any model prediction, and after any user feedback, the system takes into account all available information to appropriately update the subject-media association probability for the media item and associated subject.

Subject-media associations that have received user feedback and have sufficiently low uncertainty (below 5% by default) are declared as 'consensus' subject-media items. These consensus items are considered to be sufficiently reliable to be used for the purpose of model training. Consensus items are denoted by a green check mark for positive association and a red x for a negative association within the Cogniac web and mobile applications.

Advanced Cogniac application types (such as counting and measurement applications) can include additional information within a subject media association. For example, an image could be associated with a subject "horse" and a count of 3, which would indicate to the system that there are three horses in the image. In all cases the information is anchored by, and serves to qualify, the subject as it pertains to a particular media item.