TFIPDL

Currently, deep neural networks are the basis for the majority of artificial intelligence applications. That is due, on one hand, to the increase in the processing power of machines and, on the other hand, to their capacity for automatic learning from a suitable set of examples. Some of the neural networks used most often, especially for image processing or object identification problems, are convolutional neural networks. Essentially, those networks consist of repeating two stages: convolution, which makes it possible to identify relevant characteristics of the examples, and pooling, which makes it possible to reduce the size of the data. Both processes are usually done with the same mathematical functions, regardless of the specific problem under consideration.

In this project we intend to study potential improvements of the results of a deep neural network for object detection problems by designing specific convolution and pooling functions adapted to this specific problem. To do that we have developed a theoretical analysis of the properties that can be expected from those functions in parallel with an experimental development of a deep neural network for a specified problem. The project is the first step in identifying the most appropriate techniques for each specific problem, which opens the door to potentially significant improvements in industrial applications in the near future.

 

It can be deduced from the results obtained that the pooling modifications, except in very specific cases, do not lead to excessive improvements. On the contrary, the study suggests that convolution modifications may be highly relevant, because using functions that take relationships between characteristics of the data into account may make it possible to improve the performance of convolutional networks. Nevertheless, those changes are very complex to implement because of computational difficulties and the need to develop new mathematical techniques that can be translated into code for the networks and included in their learning mechanisms in an appropriate way.

 

At any rate, this project is only the first part of an ambitious task with which we hope to obtain completely original developments in deep neural networks that can efficiently adapt themselves to image problems that are typical in the automotive industry, like detecting components and real time image processing.


  • Año: 2019
  • Sector estratégico: Movilidad eléctrica y conectada
  • Líder del proyecto: Universidad Pública de Navarra
  • Socios del proyecto: NAITEC

2019

X