High performance scalable hardware SOM architecture for real-time vector quantization
Résumé
Real-time vector quantization requires high performances and high configurability. Unsupervised neural networks such as Self-organizing maps (SOMs) have been commonly used as vector quantizers in many application domains, mainly in software. The software counterparts of SOMs are highly flexible with limited performances and cannot be used in real-time applications where tight timing constraints require high performances. The flexibility of software SOMs and the performances of their hardware counterparts must be coupled to obtain highly configurable, scalable and application independent SOMs. In this paper, a massively parallel hardware implementation of a scalable and flexible self-organizing map (SOM) is presented. The scalability and flexibility are obtained by using highly connected communication approach tailored to SOM needs, while the distributed clusters of neuro-processors are at the origin of high performances. The proposed architecture is validated on different map sizes (up to 16×16) with different vector widths (up to 256) applied for real-time vector quantization.