AI Folk
Resource Management for Distributed AI

This project is funded by a grant of the Ministry of Research, Innovation and Digitization, CNCS - UEFISCDI, project number PN-III-P1-1.1-TE-2021-1422, within PNCDI III.

Project duration: May 2022 - June 2024

Our research goal is to develop a knowledge model and an interaction protocol which allow a system formed of multiple actors -- human, software, or organizational -- to find, use and share improvements on machine learning resources. Such resources can be datasets, models, or experiences.

We propose the development of the AI Folk framework and methodology, at the intersection of machine learning, knowledge management, and multi-agent systems. It comprises tools and methods that allow the management and discovery of ML-related resources in a distributed system. Federated learning has yet to achieve maturity as a field of study and this is a novel approach which assumes an open system and a variety of resources. We believe that this approach will lead to an advance in the state of the art and will help in the development of standards for open and distributed artificial intelligence.

The objectives for this project are to:

• develop a semantic description for the resources which can be exchanged in a distributed AI ecosystem, such as individual observations, datasets, and machine learning models.
• develop a protocol that allows actors in an AI ecosystem to search for resources, to understand how resources can be used, and to share resources that they have improved as a result of their own experience. Searches will be carried out both in a peer-to-peer manner inside the community, and in libraries of ML resources. The dimensionality of the resources may change to better suit the receiver.
• define a methodology which allows the integration of existing datasets and models for other applications into an AI Folk ecosystem.
• validate the AI Folk approach through the implementation of prototype applications in two scenarios:
  • autonomous driving, where there is a great variety of situations and a relative limitation of resources;
  • disaster response, characterized by high heterogeneity, limited resources, and unstable communication;

Papers with the project acknowledgement

• Olaru, A., Nicolae, G. & Florea, A. M. (2023). The entity-operation model for practical multi-entity deployment, in Proceedings of EMAS 2023, the 11th International Workshop on Engineering Multi-Agent Systems, in conjunction with AAMAS 2023, 29-30 May 2023, London, UK.
  paper slides
• Olaru, A. & Pricope, M. (2023). Multi-Modal Decentralized Interaction in Multi-Entity Systems in Sensors, 23(6), pp. 3139.
  paper
• Olaru, A., Sorici, A., Nan, M. & Iancu, D. (2024). AI Folk: Sharing Machine Learning Models in a Multi-Agent Community in Proceedings of DCAI'24, 21st International Conference on Distributed Computing and Artificial Intelligence, 26-28 June 2024, Salamanca, Spain.
  paper slides

Presentations

• AI Folk – Building a machine learning culture - presentation during the Management Committee meeting of Cost Action CA19134 - Distributed Knowledge Graphs (DKG), in Malaga, Spain, 26.09.2023
  slides

Deliverables

At the end of the project, we will have developed the following resources:

• a domain ontology (in RDF/S) that describes the properties of and relations between four types of resources: individual sensorial observations from devices; datasets formed of multiple observations within a specific context (e.g. images of roads in a given area); learning models, together with information about their input, output, and training data; observations regarding the experience of using a model, such as situation in which it was used, performance information, and, if possible to determine, quality of the results. The ontology will focus on autonomous driving and disaster response applications, and it will be delivered together with a methodology of expanding the ontology for other application fields.
  • the core ontology is available at aifolk-core-v2.ttl
  • the ontology for the self-driving scenario is available at aifolk-drivingsegmentation-v1.ttl
  • the ontology for the disaster response scenario is available at aifolk-disaster-response-segmentation-v1.ttl
• the specification of a high-level interaction protocol, that allows actors in the system to search for resources, transfer resources (datasets, models) locally in a manner that is efficient (using compression and incremental updates) and customized (e.g. dimensionality reduction), subscribe to updates, and share experiences with the resources (especially learning models). The protocol is implemented as a RESTful Web Service, encapsulated in a module that offers to actors an API through which they can access functionality.
• a methodology with describes how to develop the appropriate descriptions for ML resources, as well as how to construct queries for searching resources, build based on the experience with the two validation scenarios in the project.
• two prototype applications, in the field of autonomous driving and disaster response. In these applications, autonomous agents deployed in the environment will have the ability to store, search for, transfer, and use resources such as datasets or models.

The 2022 report introduces:

• a primary architecture for the entities in the framework
• a first version of the AI Folk ontology
• the design of the AI Folk interaction protocol
• FLASH-MAS as the infrastructure that will underpin interaction in an AI Folk ecosystem

Download the report here

The 2023 report presents:

• the architecture of the AI Folk ontology and its instantiation for the autonomous driving scenario
• the design, integration, and deployment of the AI Folk tools, using FLASH-MAS
• experiments with the autonomous driving scenario
• the design of the disaster response scenario

Download the report here

The final report presents:

• the achievements of the project
• description of core ontology and the instantiations for each scenario
• specification of the interaction protocol
• the architecture of the deployment platform
• the AI Folk methodology
• the two application scenarios developed

Download the report here

The project is developed by a team of young researchers ar University Politehnica of Bucharest:

Andrei Olaru, PhD, associate professor -- project coordinator
Alexandru Sorici, PhD, Lecturer
David-Traian Iancu, PhD, Lecturer
Mihai Nan, PhD, Lecturer
Teodor-Cristian Adam, Master student