The Data Types, Formats and Attributes standard

The Data Types, Formats and Attributes (MPAI-TFA) V1.5 standard specifies Qualifiers. A Qualifier is additional information attached to an instance of a Data Type, such as Text, Speech or Visual, that an AI Module (AIM) may need to process that instance properly.

The standard enables every AIM to know what the data it receives is (its form, how it is represented, and what it describes) without having to guess or be told out of band.

The Structure of a Qualifier

MPAI-TFA classifies the information in a Qualifier into three kinds:

  • Sub-Types – the different forms a Data Type instance can take; for example, Colour Space is a Sub-Type of the Visual Data Type.
  • Formats – the different ways a Data Type can be digitally represented or transported; for example, AAC is a Format of the Speech Data Type.
  • Attributes – information giving details on a Data Type instance; for example, the ID of an Object in a picture.

Sub-Types, Formats and Attributes are further organised into subordinate hierarchies. Every element of a Qualifier is optional; which elements to include is decided solely by the user.

Main Functions of Qualifiers

The MPAI-TFA standard enables the following essential functions:

  • Identification – States what a Data Type instance is before its content is processed.
  • Representation – States the digital format and transport of the data.
  • Description – Adds details about the content, such as the identity of objects it contains.
  • Selection – Lets an AIM accept, reject or adapt its processing according to the Qualifier.
  • Extension – Lets application domains request new elements of the hierarchy.

These functions allow AIMs from different developers to exchange data and interpret it in the same way.

The Domains Covered

MPAI-TFA V1.5 specifies 39 Qualifiers in eight domains:

  • Automotive – GNSS, Offline Map, Ultrasound, LiDAR and RADAR.
  • Avatar – Animation, Body Descriptors, Face Descriptors, Gesture Descriptors and MoCap.
  • Health – Behavioural Signal, Clinical Record, ECG, EEG, EHR, Health Data, Medical Imaging, Neurophysiological Signal, Omics and Physiological Signal.
  • Machine Learning – ML Model.
  • Media – 3D Model, Audio, Audio-Visual, Colour, Speech, Text and Visual.
  • Metaverse – Certificate, Contract, Currency, Discovery, Interpretation and Program.
  • Space-Time – Location and Time.
  • Theatrical – Control, Dance and Metaverse.

Operation Model

An AIM receives an instance of a Data Type together with its Qualifier:

  • The Qualifier travels with the data it qualifies.
  • The receiving AIM reads the Qualifier first: the Sub-Type and Format tell it how to decode the data, the Attributes what the data describes.
  • An AIM that produces data attaches the Qualifier that states what it produced.
  • Elements the producer did not determine are simply omitted.

This model lets the same Data Type carry data in many forms while every AIM along a workflow knows which form it is handling.

Powered by the MPAI AI Framework

MPAI-TFA Qualifiers are used by AIMs executed in the MPAI AI Framework (MPAI-AIF), which provides:

  • A modular and interoperable execution environment for AIMs.
  • Configuration and orchestration of AIMs in an AI Workflow.
  • Platform-independent implementation.

Qualifiers are referenced by the Data Types of other MPAI standards, so the same Qualifier serves every standard that uses that Data Type.

Open and Flexible Design

MPAI-TFA follows key design principles:

  • One Qualifier per Data Type, shared across all MPAI standards.
  • Optional elements. A Qualifier carries only what its producer knows.
  • Hierarchical organisation. Sub-Types, Formats and Attributes can be refined without changing their structure.
  • Domain-driven growth. New elements are added on request from application domains.

Benefits

The MPAI-TFA standard will support:

  • AIM developers to receive data whose nature is stated rather than assumed.
  • Integrators to connect AIMs from different providers without format mismatches.
  • Standard developers to reuse Qualifiers instead of redefining them in each standard.
  • Application domains, from automotive to health and theatre, to describe their data in a common way.
  • Users to obtain more reliable results from AI systems that process data correctly.

A New Paradigm for Data in AI Systems

MPAI-TFA promotes a shift from data whose meaning is implicit to:

  • Data that states its own type, form and format.
  • A common description of data across standards and domains.
  • AI components that can adapt to the data they receive.
  • A catalogue of Qualifiers that grows with the domains using it.

Conclusion

MPAI-TFA V1.5 provides an interoperable way to qualify data that:

  • States the Sub-Type, Format and Attributes of a Data Type instance.
  • Covers media, avatars, vehicles, health, machine learning, the metaverse, space-time and theatre.
  • Keeps every element optional and every hierarchy extensible.

The standard enables AI Modules to exchange data whose nature is understood the same way by all.