Learning sparse neural topologies for embedded avionic applications

Occurrences : 


Nombre d'étudiants minimum: 


Nombre d'étudiants maximum: 


Nombre d'instances : 


Convolutional Neural Networks (CNNs) can be trained to achieve state-of-the-art performance in several computer vision tasks thanks to complex topologies with millions of learnable parameters. Such complex topologies limit however the chances to deploy such architectures on embedded devices, where installed memory and computational resources are limited. Nevertheless, once a network has been trained at a task, part of the connections between neurons can be dropped, yielding a sparse network topology with reduced memory foot print. In this project, we propose to train a CNN for character recognition with sparse topology for use in an avionic application. The project requires some knowledge of neural networks and the knowledge of the pytorch framework is desirable.

References for further information: