Distributed Learning for Robust Fuzzy Neural Networks
- Publication Type:
- Thesis
- Issue Date:
- 2023
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
The world of machine learning is constantly evolving, striving to better manage and learn from increasingly complex and diverse datasets. However, two of the most pervasive issues that are yet to be effectively addressed are the challenges posed by high-dimensionality and high-uncertainty data. These difficulties are even more pronounced in multi-agent systems, where distributed fuzzy neural networks (DFNNs) are commonly employed due to their ability to process uncertainty locally. Despite their usefulness, DFNNs have limitations, such as struggling with high-dimensional data, extreme uncertainty levels, and personalizing models for non-independent and identically distributed (non-IID) local data. Moreover, they often overlook an abundance of easily-obtainable unlabelled data, which holds untapped potential for enhancing the robustness and generalizability of models.
Addressing these issues, this research proposes an innovative solution through the development of Robust Fuzzy Neural Networks (RFNNs). Unlike their traditional counterparts, these RFNNs are designed to handle both high-dimensionality and high-uncertainty data effectively. They improve local client customization by implementing an adaptive inference engine that can learn the firing strength of each fuzzy rule, regardless of the high dimensionality of the input data. Furthermore, they assign an uncertainty tolerance to membership values, effectively tackling data with high uncertainty. The models also incorporate neural-structured consequent components, which leverage neural network structures to enhance the reasoning ability of fuzzy rules when dealing with complex inputs.
To better utilize the vast amount of unlabelled data available, this research also introduces an RFNN enhanced with Interpolation Consistency Regularisation (ICR). This technique was recently proposed to regularise semi-supervised problems and can enforce decision boundaries to pass through sparse data areas, thus increasing a model's robustness. In tandem with this, a novel distributed semi-supervised fuzzy regression (DSFR) learning method is proposed. The DSFR model introduces a novel distributed fuzzy C-means method and a distributed variant of interpolation consistency regularisation.
Lastly, considering the paramount importance of privacy preservation in today's data-driven world, this research presents a Federated Fuzzy Neural Network (FedFNN) with evolutionary rule learning (ERL) for decentralized scenarios. The FedFNN approach manages data heterogeneity, personalized privacy preservation, and high levels of data uncertainty by maintaining a global set of rules in the central server and a personalized subset of these rules for each local client. The ERL approach, inspired by biological evolution, encourages rule variations while activating superior rules and deactivating inferior ones, thus effectively handling non-IID data.
Please use this identifier to cite or link to this item: