The AIW and AIM Implementation Guidelines project

The Technical Report AIW and AIM Implementation Guidelines (MPAI-WMG) V1.0 accompanies six MPAI Technical Specifications. It examines all the AI Workflows (AIWs) they specify and a subset of their AI Modules (AIMs), to help those who implement them.

The Report clarifies the specifications where appropriate, analyses the issues their implementation raises, and studies how they fit MPAI’s model of Perceptible and Agentive AI (PAAI).

The Specifications Covered

MPAI-WMG V1.0 considers the AIWs and AIMs of:

  • Connected Autonomous Vehicle (MPAI-CAV) – Technologies (CAV-TEC) V1.0.
  • Context-based Audio Enhancement (MPAI-CAE) – Use Cases (CAE-USC) V2.3.
  • Human and Machine Communication (MPAI-HMC) V2.0.
  • Multimodal Conversation (MPAI-MMC) V2.3.
  • Object and Scene Description (MPAI-OSD) V1.3.
  • Portable Avatar Format (MPAI-PAF) V1.4.

Each specification has its own section, organised as its AI Workflows followed by its AI Modules.

The Perceptible and Agentive AI Model

A PAAI is a machine, or a set of machines, that may:

  • Perceive a physical or virtual environment, including objects, humans and other PAAIs, and access information sources.
  • Represent, describe and interpret data from objects and messages from humans, and receive messages from PAAIs.
  • Reason about data and messages.
  • Set goals, decide on messages, actuate decisions, store and retrieve experiences, and adjust goals.
  • Make plans: decompose tasks into subtasks, consider the environment and resources, involve processes, humans and PAAIs, and iterate.
  • Execute the plan, changing it when conditions change, and learn while experiencing.

To do so, a PAAI uses twelve elementary functionalities:

  • Representation – captured audio-visual information as Data.
  • Description – Data as Descriptors, e.g. AV Scene Descriptors.
  • Interpretation – Descriptors as Interpretations, e.g. speech recognition.
  • Conclusion – the result of reasoning about Interpretations.
  • Communication – exchange of Data with another PAAI.
  • Goal setting – the goals to be reached.
  • Planning – structured plans to reach a goal.
  • Decision – how to implement a Conclusion.
  • Explanation – the path that led to a Decision.
  • Action – actuation of a Decision.
  • Storage/Retrieval – of Experiences, the relevant data a PAAI produces during its operation.
  • Learning – improvement of a stage while experiencing it.

How the Report Analyses a Specification

MPAI-WMG treats every AIW as a PAAI composed of collaborating PAAIs, its AIMs.

For each AIW, the Report:

  • Lists its AIMs and the role each plays in the workflow.
  • Gives the Reference Model of the workflow.
  • States the level of operations the workflow performs, such as the Descriptors or the Interpretation level.

For each AIM analysed, the Report:

  • States what the AIM receives, what it transforms or finds, and what it produces.
  • Notes implementation requirements, e.g. graphic rendering capabilities, or the possibility of using neural networks.
  • States the level of operations the AIM performs.

For example, the Avatar Videoconference Server of MPAI-PAF is analysed as four collaborating PAAIs: Portable Avatar Demultiplexing, Text and Speech Translation, Service Participant Authentication and Portable Avatar Multiplexing. It performs Interpretation Level Operations.

Powered by the MPAI AI Framework

The AIWs and AIMs covered by MPAI-WMG are specified to run 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.

The guidance in the Report therefore applies to any AIF-based implementation of these specifications.

Open and Flexible Design

MPAI-WMG follows key principles:

  • One model for all. The same PAAI model applies to vehicles, audio, communication, conversation, scene description and avatars.
  • Uniform analysis. Every AIM is described by what it receives, what it does and what it produces.
  • Levels made explicit. Each AIW and AIM states whether it works on Descriptors or on Interpretations.
  • Guidance, not prescription. The Report clarifies and analyses the specifications it covers.

Benefits

MPAI-WMG will support:

  • Implementers to build AIWs and AIMs with a clearer understanding of what each must do.
  • Developers of AIMs to see how their module fits the workflows that use it.
  • Integrators to combine AIMs from different specifications on a common model.
  • Standard developers to identify where specifications need clarification.
  • Users to rely on implementations that behave consistently.

A New Paradigm for AI Implementation

MPAI-WMG promotes a shift from reading each specification in isolation to:

  • A common model of perceiving, reasoning and acting across all MPAI specifications.
  • AI systems understood as collaborating agents, each with a defined role.
  • Implementation guided by the level at which each component operates.
  • A shared vocabulary between specification writers and implementers.

Conclusion

MPAI-WMG V1.0 provides a common guide to the implementation of MPAI AI Workflows and AI Modules that:

  • Clarifies the specifications where needed.
  • Analyses the implementation of all AIWs and a subset of their AIMs.
  • Applies a single model, Perceptible and Agentive AI, across six specifications.

The Report enables more consistent and interoperable implementations of MPAI standards.