Projecte llegit
Títol: Reducció del cost computacional en l'ús de xarxes neuronals del Cloud a l'Edge per aplicacions en temps real.
Director/a: BERTRAN ALBERTÍ, EDUARD
Departament: TSC
Títol: Reducció del cost computacional en l'ús de xarxes neuronals del Cloud a l'Edge per aplicacions en temps real.
Data inici oferta: 27-01-2021 Data finalització oferta: 27-09-2021
Estudis d'assignació del projecte:
DG ENG SISTE/TELEMÀT
| Tipus: Individual | |
| Lloc de realització: Fora UPC | |
| Supervisor/a extern: Miguel Reyes Estany | |
| Institució/Empresa: IDNEO TECHNOLOGIES S.A.U | |
| Titulació del Director/a: Enginyer en Informática/ Doctor matemati | |
| Paraules clau: | |
| Deep learning, xarxes neuronals, optimització, Edge, quantització, pruning. | |
| Descripció del contingut i pla d'activitats: | |
| L¿empresa IDNEO està desenvolupant una plataforma de Deep Learning al Cloud anomenada SABIA. En moltes aplicacions on s¿utilitza el Deep Learning la latència és un aspecte fonamental a tenir en compte i, per tal de reduir-lo, es porta la lògica neuronal al dispositiu final (edge). La necessitat d¿optimitzar ve donada pel fet de que al Cloud els recursos computacionals són il·limitats, per tant, si volem processar les dades directament al dispositiu final, serà necessari reduir la càrrega del model de xarxa neuronal utilitzat. L¿objectiu d¿aquest treball és l¿estudi i l¿aplicació de tècniques d¿optimització de xarxes neuronals per a ser utilitzades a dispositius a l¿Edge. | |
| Overview (resum en anglès): | |
|
Internet of Things is currently taking on a very important role in society, which is becoming increasingly connected. In recent years, an exponential increase in devices has occurred and, consequently, so has the volume of data generated every day. This increase in data has led to the ability to develop more complex artificial intelligence models to work with artificial neural networks. More and more applications working with this technology require an immediate response. Consequently, for the applications that require low latency, is no longer so beneficial to work with the unlimited resources provided by the cloud, but rather the need to bring the processing tasks to the device itself: the edge. Bringing cloud's inference of complex models to edge brings a number of drawbacks, most significantly the lack of computational power. The resources that are obtained from a device that a conventional user can use, are potentially lower than those we would get from a centralized system. So it is necessary to reduce the computational cost of processing an artificial neural network. There are currently several techniques for optimizing models, but there is still no clear trend and a single solution. For this reason, it has been considered interesting to analyse the two most widely used and advanced techniques at the moment: quantization and pruning. To achieve this, a theoretical explanation of the concepts necessary to understand them and practical examples using PyTorch have been documented. Finally, the development and optimization of an application capable of detecting Melanomas in the skin has been carried out. |
|