<–Towards a responsible AI Some MPAI data coding standards–>

The development of a market of AI solutions has to take into account the following basic components:

  • All players (e.g., Microsoft, Amazon, Google) provide environments supporting the AI application life cycle with their frameworks (e.g., Azure AI Platform, Sagemaker, Greengrass, TensorFlow etc.) and offer a store.

  • In general, migration of an application to a different environment is complicated because exporting models is not easy.

  • Typically, an AI application requires large and scarce multidisciplinary competences that span from data science to domain knowledge and consequently require large capital investments in both human and computational resources.

  • Models and guidelines for the development of explainable AI applications are in their infancy. Current applications are monolithic and opaque making their adoption at scale problematic.

There is no overriding reason, however, for AI applications to be monolithic. If we look at the human brain, for example, we do not find a single network but a collection of connected specialised subnetworks (and sub-subnetworks).

Even at the current state of brain research, it is possible to identify and characterise the functions of subsystems of the human brain.

Whether inspired or not by this reference, in its standards MPAI divides AI applications serving identified use cases into units called AI Modules (AIM) that are defined by their functions, interfaces and input/output data. Thanks to the way they are defined, AIMs can be combined in a way that is agnostic of the AIM provider, on condition that the component has been developed according to the standard. Additionally, AIMs can be combined to address scenarios that the AIM provider may not even have foreseen. Because AI is quite a different technology than those of the past, verifying standard conformance requires steps that enable MPAI implementation users to make informed decision about their applicability. Central to this are the notions of Conformance and Performance, the latter defined as a set of attributes characterising a reliable and trustworthy implementation.

Thus, a full AI application is obtained by interconnecting the selected AIMs to create a “workflow” of AIMs, that MPAI calls AI Workflow (AIW). Again, an AIW is defined by its function as identified by the use case, its interfaces, and the format of the input/output data.

The MPAI-specified framework – called AI Framework (AIF) – allows an application implementer to develop ready-to-use systems that combine AIMs. These may:

  1. Have been developed in any of the mentioned environments using any of proprietary frameworks for any operating system.

  2. Be AI-based and non-AI-based.

  3. Be implemented in hardware or software or in a hybrid hardware and software combination.

  4. Execute in Microcontroller Units (MCU) to High Performance Computers (HPC) in local and distributed environments, and in proximity with other AIFs.

irrespective of the AIM provider. Of course, appropriate profiles of the AIF standard may need to be defined to enable such a wide range of application contexts.

Figure 14 depicts the framework components of the MPAI approach to AI data coding standardisation:

  1. AIF, AIW and AIM, as described above.

  2. Controller, a computing component that exposes Application Programming Interfaces (API) to the AIMs and to the

  3. User Agent, the means by which a user acts with the system.

  4. Two data storage components, one specific to an AIM and another accessible by all AIMs).

  5. Access to external slowly varying data.

  6. MPAI Store, a platform containing and managing a repository of AIM, AIW and AIF Implementations managed by a non-profit commercial organisation established and controlled by MPAI.

Figure 14 – The AI Framework (AIF) reference model and its components

The MPAI AI Framework offers several key advantages:

  1. Component providers can offer conforming AIMs to an open competitive market.

  2. Application developers can find the AIMs they need on the open competitive market.

  3. Consumers have a wider choice of better AI applications from competing developers.

  4. Innovation is fuelled by the demand for novel/more performing AIMs because constraints imposed by proprietary interfaces are removed and the end user’s choice exclusively depends on the quality of the implementation.

  5. Society can lift the veil of opacity from large, monolithic AI-based applications.

The operation of the Store is well described by the role played by the following actors:

  1. Implementers:

    1. Upload their implementations to the Store.

    2. Submit their implementations to a Performance Assessor (if required).

  2. MPAI Store:

    1. Verifies the implementation for security.

    2. Tests the implementation for conformance, i.e., that the it is a correct implementation of the standard providing a minimum level of functionality.

    3. If required, checks that the submitted implementation has been assessed by a Performance Assessors.

    4. Makes implementations available declaring their Performance grade.

  3. Users:

    1. Pay fees on a cost-recovery basis.

    2. Download implementations.

    3. Report scores of user experience to the MPAI Store.

The MPAI Store has several desirable features for both implementers and end users:

  1. It supports a market for both components (AIM) and applications (AIW).

  2. It has an implementer-friendly business model like today’s app market.

  3. It promotes competition because different AIMs and AIWs with the same function and interfaces can be posted to the Store.

  4. It offers implementations with different levels of AIW interoperability:

    1. Level 1 – The AIW is implementer-specific but conforms with the MPAI-AIF Standard.

    2. Level 2 – The AIW conforms with a use case specified by an MPAI application standard.

    3. Level 3 – The AIW conforms with a use case specified by an MPAI application standard and its AIMs are certified by Performance Assessors.

The ecosystem created by MPAI has several desirable features:

  1. Application developers can build diverse applications (AIW) because AIMs can be integrated in creative ways.

  2. AIMs can individually be of high quality because implementers can post implementations issued from a specialised field to the Store.

  3. AI technologies can develop faster and better because the market is competitive at the level of its basic units (AIM).

The ecosystem described above will need governance. Chapter 17 defines how the MPAI ecosystem will be governed.

<–Towards a responsible AI Some MPAI data coding standards–>