Titulo Estágio
GPU Implementation of Radial Basis Functions Networks
Área Tecnológica
GPU Computing, Adaptive Computation, Distributed Systems,
Local do Estágio
Centro de Informática e Sistemas, CISUC, Grupo de Computação Adaptativa, Lab: LARN
Enquadramento
The Graphics Processing Unit (GPU) has become an integral part of today’s
mainstream computing systems. The enormous computational potential of GPUs has
led to research for general-purpose computation on GPUs (GPGPU). It uses GPU
for non-graphics computations such as image processing, neural networks (NNs),
linear algebra, sorting, computational physics and database queries. Over the past few years, the raw computational power of GPUs due to its
parallelism has surpassed by far that of top range CPUs. Unlike general-purpose
processors, GPUs are optimized to perform floating-point operations on large
data sets using the paradigm Single Instruction Multiple Data (SIMD). This is special important with machine
learning algorithms (such as neural networks, kernel machines, etc.) which are
often complex, placing high demands on memory and computing resources and CPUs are simply not powerful enough to solve
them quickly for use in interactive applications. However, this architectural
difference leads to more complex programming tasks since CPUs use the paradigm
SISD (Single Instruction Single Data). To cope with this
complexity NVIDIA developed a parallel technology namely CUDA (Compute United
Device Architecture) which provides a programming model for its GPUs
with an adequate API for non-graphics applications using standard ANSI C,
extended with keywords that designate data-parallel functions.
Objetivo
The goal of this thesis is to design, develop and
implement a component to integrate/support GPUNetLib
software which aims to provide users with a Library of neural network tools
taking advantage of their very fast implementation in the GPUs. One of the main limitations of NNs is the long time required for the
training process. In particular, radial basis function network (RBF) has been
used extensively, proving its power. However, their limiting time performances
reduce their application in many areas.
A possible way to accelerate the learning process of an NN is to
implement it in hardware, but due to the high cost and the reduced flexibility
of the original central processing unit (CPU) implementation, this solution is
often not chosen. The power of the graphic processing unit (GPU), on the
market, has boosted its use for implementation of data dependant algorithms
such as NNs. A GPU implementation of the entire learning process of an RBF
network should reduce the computational cost making them algorithms of choice
for many applications. Global Thesis Planning
Plano de Trabalhos - Semestre 1
In this phase, the candidate should make a
review of the state-of-the-art of the technology and should make a report with existing
algorithms implementing neural networks in particular RBF (Radial Basis
Functions).
In this task, the candidate will implement an NN
algorithm using CUDA. The initial goal of this phase is to allow to the student
to acquire core competences in this technology. Besides, the candidate should
acquire a concise notion of the performance of GPU to optimize the work in next
phase.
As a result of the previous phase and prior to
Thesis Proposal writing, functionalities to be included in the RBF component
should be defined.
Plano de Trabalhos - Semestre 2
RBF
Component Implementation of GPUNetLib (months 6-9). The implementation code for the RBF software component should be written Thesis Final Report (month 10)
Condições
The work will be in the Departament of Informatics
Engineering, Adaptive Computation Group, Laboratory of Neural Networks
(LARN) of the Center for Informatics and
Systems of the No grant is available for this internship. MEI- Mestrado em Eng. Informática – DEI/FCTUC
Observações
Orientador
Bernardete Ribeiro & Noel Lopes
bribeiro@dei.uc.pt 📩