Hyperdimensional Computing for Efficient Distributed Classification with Randomized Neural Networks

The recent evolution of artificial intelligence has made a structural limitation of traditional models increasingly evident: their dependence on centralised data. In many real-world scenarios, data cannot be gathered in a single location nor freely shared across systems. In this context, distributed learning becomes not just an option but a necessity, introducing new challenges related to cooperation among models and communication costs. The work develops a solution that directly addresses this issue by proposing a distributed classification framework based on randomised neural networks and a compression mechanism derived from Hyperdimensional Computing. The objective is not only to learn from distributed data, but to do so efficiently, significantly reducing the amount of information exchanged among agents.

The starting point is a scenario in which multiple agents operate on different portions of data, without direct access to each other’s datasets. Each node builds its own local model using only the available information. However, these models do not remain isolated: each agent shares a representation of its classifier with others, contributing to the construction of collective knowledge. Unlike centralised or orchestrated approaches, there is no coordinating node. The entire process unfolds in a fully distributed manner, where each agent autonomously integrates the received information, combining local models into a more effective aggregated version.

The architecture is based on Random Vector Functional Link networks, a class of neural networks where most connections are randomly generated and kept fixed. This approach significantly reduces training complexity, focusing learning only on the final part of the model, namely the classifier. To represent information, the Hyperdimensional Computing paradigm is introduced, which uses high-dimensional vectors to encode data and relationships. Through operations such as binding and superposition, it becomes possible to construct distributed representations that preserve relevant information even in the presence of noise or partial data loss.

The most innovative aspect lies in how local models are shared. Instead of transmitting the classifier directly, each agent compresses it into a single high-dimensional hypervector. This representation is obtained by associating each class with a random key and combining information through circular convolution and addition operations. The result is a lossy compression scheme, where part of the original information is inevitably lost. However, this loss does not compromise the system’s effectiveness. On the contrary, when multiple compressed models are combined, the noise tends to average out, allowing the reconstruction of an effective approximation of the global classifier. Compression thus becomes a central design element, enabling reduced communication overhead while maintaining competitive accuracy compared to traditional approaches.

The experimental validation reveals a particularly significant outcome. Local models, when considered individually, perform worse than a centralised model, especially as the number of agents increases and the amount of data per node decreases. However, when agents share and aggregate their classifiers, performance improves substantially. The distributed version often approaches the performance of the centralised model, despite having no access to the full dataset. In some cases, the use of compression even introduces a form of regularisation, further improving the model’s generalisation capabilities.

The most important contribution goes beyond efficient distributed classification. It offers a broader perspective on the role of artificial intelligence in decentralised environments. The work demonstrates that it is possible to build systems capable of collective learning without sharing sensitive data, while also reducing communication costs. In this view, Hyperdimensional Computing is not just an alternative representation technique, but a practical tool for designing cooperative models. Intelligence no longer resides in a single system, but emerges from the interaction of multiple agents, each with a partial view of the problem. When communication becomes as critical as computation, the quality of a model is measured not only by how well it learns, but by how effectively it collaborates under real-world constraints.