Titulo Estágio
Self Organizing Maps (SOMs) in Graphics Processing Units for Object Recognition
Área Tecnológica
Reconhecimento de Padrões
Local do Estágio
DEI - Departamento de Engenharia Informática
Enquadramento
The Graphics Processing Unit (GPU) has become an integral part of today's mainstream computing systems. GPUs are optimized to process large quantities of data using the paradigm Single Instruction Multiple Data (SIMD). To this end, a typical GPU possesses a much larger number of processing cores than CPUs (for example the new NVIDIA GTX 680 contains 1536 cores while an Intel core i7 CPU may have up to 6 cores). Hence, recently the GPU emerged as a major player in the trend shifting of the computing industry from single core to multi-core architectures with tremendous gains for a range of applications in many areas and in particular in the machine learning area. Machine learning algorithms are among those for which a GPU implementation can deliver a significant performance impact, because they are often complex, placing high demands on typical computing resources.
However, this architectural difference (between GPUs and CPUs) leads to more complex programming tasks. To simplify this task, NVIDIA developed CUDA (Compute United Device Architecture) which provides architecture and a programming model for developing general purpose applications on the GPU (GPGPU) using C/C++, extended with keywords that designate data-parallel functions.
Objetivo
In this context, the goal of this thesis is to design, develop and implement a component (Kohonen’s Self-Organizing Map (SOM)) to integrate/support GPUMLib (http://gpumlib.sourceforge.net/) - an open source Graphic Processing Unit Machine Learning Library - which aims to provide users with a Library of machine learning tools taking advantage of their fast implementation in the GPUs. More specifically, the component to be developed includes the implementation, test and experimentation. Tests will be held in an object recognition problem in order to compare both the performances in the GPU and in an SOM standalone CPU version.
Plano de Trabalhos - Semestre 1
1) Introduction to CUDA, Sept 2013
2) Review of the state-of-the-art Sep-Oct 2013
3) Analysis and Specification of SOMs in the GPUMlib Nov-Dec 2013
4) Intermediate report and plan of possible extensions or alternatives: Jan 2013
Plano de Trabalhos - Semestre 2
1) Development and integration of the software component in GPUMLib Feb-May 2013
2) Tests and Experimentation on machine learning databases Mar-May 2013
3) Thesis Report May-Jun 2013
Condições
Strong skills in programming (Java, Python, C/C++).
Other interesting (optional) skills/interests include Machine Learning and Pattern Recognition
Observações
The candidate curriculum is required.
Orientador
Bernardete Ribeiro e Noel Lopes
bribeiro@dei.uc.pt 📩