NP-SOM: Network Programmable Self-Organizing Maps

Abstract

Self-organizing maps (SOM) are a wellknown and biologically plausible model of input-driven selforganization that has shown to be effective in a wide range of applications. We want to use SOMs to control the processing cores of a massively parallel digital reconfigurable hardware, taking into account the communication constraints of its underlying network-on-chip (NoC) thanks to bio-inspired principles of structural plasticity. Although the SOM accounts for synaptic plasticity, it doesn't address structural plasticity. Therefore we have developed a model, namely the NP-SOM (network programmable self-organizing map), able to define SOMs with different underlying topologies as the result of a specific configuration of the associated NoC. To gain insights on a future introduction of advanced structural plasticity rules that will induce dynamic topological modifications, we investigate and quantify the effects of different hardwarecompatible topologies on the SOM performance. To perform our tests we consider a lossy image compression as an illustrative application.

Cellular Self-Organising Maps - CSOM

“…In order to evaluate how CSOM behave in real-world applications, we have decided to train them on data extracted from image sequences. Self-organising maps can be interestingly applied to lossy image compression [8][9][10]4]. The principle is to split the image into non overlapping thumbnails, then learn a good quantisation of these thumbnails.…”

“…As an example, we consider sequences of 384 × 288 images subdivided into 6912 thumbnails of 4 × 4 pixels. Using a 8 × 8 SOM or CSOM (b = 6), the compression ratio already reaches a very high value of 7 before even further compressing the index list using differential and entropy coding (see [10] for details on computing the exact compression ratio).…”