Paralelizace neuronové sítě Growing Neural Gas

Abstract

This thesis investigates the parallelization of the Growing Neural Gas (GNG) algorithm, an incremental self-organizing neural network used for vector quantization and topological mapping. Six variants of the algorithm were designed and implemented in C# on the .NET 8 platform: a sequential reference implementation, two CPU-parallel variants based on the Task Parallel Library and ThreadPool, and three GPU-accelerated variants built on ManagedCUDA and ILGPU with different kernel definition strategies. The variants were experimentally compared on the C4 dataset from the RedPajama-Data-1T collection using hash-based vectorization. Experiments showed that GPU acceleration provides substantial speedup over the sequential baseline while maintaining comparable clustering quality. CPU parallelism, however, offered limited benefit at the given network size. Inherent limitations arising from the online nature of the GNG algorithm were also identified and quantified.