Cogniac Subjects

A subject is the central means by which media is grouped and managed within the Cogniac system. Subjects are fairly arbitrary and are defined by the users of the system. Subjects are generally related to the goals of the visual observation task that is being automated. In the simplest form a subject can be thought of as a ‘tag’ that can be associated with an image. In the fullest form a subject represents any user-defined concept that can be present in, or associated with, the content or domain of a group of images or video. Subjects can be simple concepts such as “cat” or can represent more complex information such as “cat with mouse in mouth”. A subject can also be the “domain” of an image generation process, such as “cat door camera”, or any such logical grouping of related images.

Subjects as Application Inputs and Outputs

Media flows through applications via Cogniac subjects. Subjects define both the input and output of applications.

It is helpful to create subjects with appropriately specific subject names to save confusion from subjects in use by other users within the same tenant. For example: 'lobby security camera feed' is a good subject name for video or images captured from from a security camera.

However subject names need not be unique, and concise output subject names are helpful during application training. For example: 'security alert', 'visitor', and 'employee' are appropriate output subject names for a security camera classification application.

🚧

Subject UID

Subject names are not required to be unique within the Cogniac System. Upon creation of a subject, a Subject UID is assigned that uniquely identifies that subject throughout the system.

Unique Subject UIDs allow:

1. Subject names and meta-data can be edited without losing any pre-existing subject-media associations.

2. Several subjects can exist with the same subject name, yet refer to different items of interest, without polluting subject-media associations between the subjects.

Subject-Media Associations

The main purpose of a Cogniac application is to make associations between media items and Cogniac subjects, effectively 'labeling' the media with the subject.

The strength of a subject-media association is defined by a probability of that subject being associated with the media item. Association probabilities close to 1.0 denote a very strong positive association; e.g. "this image contains a cat," and a probabilities close to 0.0 denote a very strong negative association; e.g. "this image does not contain a cat."

Applications based on deep convolutional neural networks learn to 'route' images from the application's input subject(s) to the correct output subjects with an appropriate probability. As with all supervised learning systems this learning process requires correctly labeled example images. Application users provide feedback to the Cogniac system on past model predictions which enables the system to create the 'consensus' subject media associations which provide the basis for model training.

Both application models and user feedback are calibrated by the Cogniac system and automatically assessed for accuracy. An application's predictions become stronger as its model accuracy improves based on accurate user feedback.

Subject-media associations can also be made through direct upload of media to a subject. This creates an authoritatively strong subject-media association. This is usually used for inputting media into the system for processing by associating media with an input subject. Alternatively, when media is directly uploaded to an output subject it is equivalent to directly 'labeling' the media with the subject for training purposes. Care must be taken when directly associating media with output subjects since this upload mechanisms bypasses the feedback consensus system that normally ensures the accuracy of the consensus labels in output subjects.