Cognitive engines can be trained to identify and process specific entities in a library. Libraries can be private to an organization and shared between organizations.
When an engine runs in training mode, it consumes identifier assets from the provided library and generates a training model. This model is the engine's internal representation of the library and an engine state that can be referenced.
Veritone's Library Application enables collections of digital assets to be created for custom engine training.
Terminology
Library- A named collection of entities an organization is interested in identifying in media, e.g. American Politicians. A library's type defines what type of entities it can hold.
Entity- An aggregation of assets (entity identifiers) for a defined concept, such as a person, company or organization, advertising campaign, or type of object.
Entity identifier- Assets associated with an entity, for example headshots for face recognition, voice clips for speaker recognition, ad creative for audio fingerprinting, DLM for transcription, or aliases for transcription.
Entity identifier type- The type of asset associated with an entity. For example, a headshot, logo, voice clip, ad, or DLM. The library type defines what identifier types it can support.
Library type- The type of entities the library contains. Used to tailor the UI experience for specific types of entities, such as people, ads, etc.
Library version- A numerically-indexed revision of a library, containing one or more changes to its entities or assets.
Shared library- A library that is shared from one owner organization to one or more collaborators or organizations.
Training- The required step an engine takes to convert the user's library into a format the engine understands.
Training model- Data generated during an engine's training step. The model is (in some cases) provided to an engine when the engine is run. A model can optionally contain an asset representing the model data.
Cross-file identification- Identification of the same entity in a library when run on multiple media files containing that entity.
Cross-engine identification- Identification of the same entity by different engines trained using the same library.
Object relationship diagram
