GPGPU implementation of growing neural gas: Application to 3D scene reconstruction

Abstract

Self-organising neural models have the ability to provide a good representation of the input space. In particular the Growing Neural Gas (GNG) is a suitable model because of its flexibility, rapid adaptation and excellent quality of representation.However, this type of learning is time consuming, specially for high-dimensional input data. Since real applications often work under time constraints, it is necessary to adapt the learning process in order to complete it in a predefined time.This paper proposes a Graphics Processing Unit (GPU) parallel implementation of the GNG with Compute Unified Device Architecture (CUDA). In contrast to existing algorithms, the proposed GPU implementation allows the acceleration of the learning process keeping a good quality of representation. Comparative experiments using iterative, parallel and hybrid implementation are carried out to demonstrate the effectiveness of CUDA implementation. The results show that GNG learning with proposed implementation achieves a speed-up of 6x comEmail addresses: sorts@dtic.ua.es (Sergio Orts), jgarcia@dtic.ua.es (Jose Garcia-Rodriguez), dviejo@dccia.ua.es (Diego Viejo), miguel.cazorla@ua.es (Miguel Cazorla), vmorell@dccia.ua.es (Vicente Morell) Preprint submitted to Parallel and Distributed Computing May 9, 2012 pared with single-threaded CPU implementation. GPU implementation has been also applied to a real application with time constraints: acceleration of 3D scene reconstruction for egomotion, in order to validate the proposal.

Comparing parallel hardware architectures for visually guided robot navigation

“…However, our results confirm that the common comparison between serial CPU implementations and GPU implementations is quite misleading (e.g., ): In such scenarios, very large speedups in favor of GPUs are achieved; however, the GPU advantage either disappears completely as soon as the full CPU capacities are utilized or at least becomes considerably smaller (speedups in the single–digit range). This is what we observe for the min‐warping algorithm as well with speedups between 2 and 8.4 in favor of GPUs compared with multi‐core‐SIMD.…”

“…There exists a considerable amount of studies in which the performance of CPU and GPU implementations is compared for specific tasks (for example, ). Even closer to our work are studies which include FPGAs or directly compare FPGAs with GPUs .…”

“…Even closer to our work are studies which include FPGAs or directly compare FPGAs with GPUs . The benchmarked applications are from many different fields like machine learning , neural modeling , optimization , numerical algorithms , image and video processing [16, 22, 24–27, 30], computer tomography , molecular sequencing , financial simulations , encryption and decoding , or analog circuit simulation .…”

3D maps representation using GNG

“…However, we designed in a previous work [17] a GPU-based implementation of the GNG algorithm that speeds up the sequential version several times. The speedup is increased as the number of neurons used for the representation grows.…”

3D model reconstruction using neural gas accelerated on GPU

“…Moreover, the NG learning algorithm has a higher level of parallelism and is more suitable to be implemented onto graphic processor units (GPUs) compared to other growing self-organising networks which have been already implemented onto the GPU [27,28].…”