Holonic Learning: A Flexible Agent-based Distributed Machine Learning Framework

TL;DR

Holonic Learning (HoL) tackles distributed machine learning with privacy and heterogeneity by organizing autonomous learning agents into a self-similar holarchy. The paper presents Holonic Averaging Learning (HoAL) as a concrete, weighted-averaging instantiation and shows how it encompasses standard distributed paradigms such as Federated Averaging, Peer-to-Peer Learning, and Hierarchical Federated Learning as special cases. Empirical results on MNIST under IID and non-IID data demonstrate convergence and competitive performance, while highlighting holarchy-dependent effects on generalization and training dynamics. The framework offers a scalable, privacy-conscious approach to distributed learning, with potential for large-scale, heterogeneous holarchies and policy-driven inter-holon interactions.

Abstract

Ever-increasing ubiquity of data and computational resources in the last decade have propelled a notable transition in the machine learning paradigm towards more distributed approaches. Such a transition seeks to not only tackle the scalability and resource distribution challenges but also to address pressing privacy and security concerns. To contribute to the ongoing discourse, this paper introduces Holonic Learning (HoL), a collaborative and privacy-focused learning framework designed for training deep learning models. By leveraging holonic concepts, the HoL framework establishes a structured self-similar hierarchy in the learning process, enabling more nuanced control over collaborations through the individual model aggregation approach of each holon, along with their intra-holon commitment and communication patterns. HoL, in its general form, provides extensive design and flexibility potentials. For empirical analysis and to demonstrate its effectiveness, this paper implements HoloAvg, a special variant of HoL that employs weighted averaging for model aggregation across all holons. The convergence of the proposed method is validated through experiments on both IID and Non-IID settings of the standard MNISt dataset. Furthermore, the performance behaviors of HoL are investigated under various holarchical designs and data distribution scenarios. The presented results affirm HoL's prowess in delivering competitive performance particularly, in the context of the Non-IID data distribution.

Figures (8)

• Figure 1: An example HoL model built over an IoT network
• Figure 2: An example holarchical network. The levels are numbered from top to bottom, and different color shades represent the heterogeneity of the holons.
• Figure 3: An example demonstrating the flow of communications in Algorithm \ref{['alg:HL_model_update']}.
• Figure 4: The holonic averaging learning model together with different special cases to implement distributed models of federated averaging, hierarchical federated averaging, and peer-to-peer models. Different shades in the lower half of the terminal holons represent the difference in their local data.
• Figure 5: The different holarchical HoAL structures used for experiments.