PUBBLICAZIONI
Optimized RVFL neural networks enable accurate AI on low-power hardware using finite precision and genetic algorithms.
The Challenge of Efficient AI in IoT Devices
In the hyper-connected world of the Internet of Things and sensor networks, many scenarios share a common challenge: processing massive amounts of data on low-cost devices with minimal computational resources. This is where an innovative solution comes into play: implementing Random Vector Functional-Link (RVFL) neural networks optimized for low-precision hardware.
Transforming Neural Networks for Limited Hardware
The key to the approach is transforming traditional RVFLs into finite precision versions, where every parameter and operation is represented with just a few bits. This is essential for microcontrollers or embedded devices that need to process data in real-time without the computational power or memory of a traditional computer.
Optimizing Precision with Genetic Algorithms
But the process goes beyond simple rounding: to avoid the dramatic drop in accuracy typical of quantization, a genetic algorithm-based optimization strategy is employed. These algorithms search for the ideal parameters for each model, significantly improving performance even with representations using just 4 or 8 bits.
Proven Performance on Real-World Datasets
The technique was tested on several real-world datasets (from aerodynamic noise to concrete strength, building energy performance, and Istanbul Stock Exchange data), showing that even with a low number of bits, results can approach those of 64-bit models, thanks to genetic optimization.
Towards Affordable and Ubiquitous Artificial Intelligence
This approach makes RVFL neural networks truly practical for distributed, low-cost scenarios, paving the way for pervasive AI that can run efficiently on affordable, energy-saving devices without sacrificing predictive accuracy. A breakthrough for those aiming to bring AI everywhere, from sensors to global networks.
Autori
A. Rosato, R. Altilio, M. Panella
