DSTN Scholars
Affiliated project Deep4EnvMonitoring
Affiliated CEA Université Gaston Berger / MITIC & Université d'Abomey-Calavi / SMIA
Title Evolutionary hyperlattice for convolutional neural networks (CNN): application to environmental monitoring
YOUME Ousmane
Affiliated project Deep4EnvMonitoring
Affiliated CEA Université Gaston Berger / MITIC & Université d'Abomey-Calavi / SMIA
Title Evolutionary hyperlattice for convolutional neural networks (CNN): application to environmental monitoring
Start date : 01/01/2021
Problematic:
Evolutionary algorithms have long been used in the field of artificial intelligence, in particular neural networks: Neuroevolution. Neuroevolution generates artificial neural networks, parameters, topology and certain rules. It is most often applied to games, evolutionary robotics, etc. It enables the evolution of a population of neural networks doing the same task, while choosing the best network over generations.
Research has shown that evolutionary neurons can play a major role in the development of visual systems. However, the application of evolutionary algorithms to convolutional neural networks has not yet been sufficiently demonstrated, due to their architecture. The aim of this topic is to combine Deep Learning, in particular CNN, evolutionary algorithms and new learning techniques such as self-supervising to generate an evolutionary model with less training data and greater recognition and adaptation range.
We apply it in the context of remote sensing (using drones or satellites) of environmental alterations observed in emerging countries, such as illegal waste dumps, precarious or unfinished housing, anarchic land occupation, etc. We also apply it in the context of remote sensing (using drones or satellites) of environmental alterations observed in emerging countries, such as illegal waste dumps, precarious or unfinished housing, anarchic land occupation, etc.
Objectives / Expected results :
The overall aim of this thesis project is to demonstrate the possibility of indirect encoding of the weights of convolutional neural networks made up of thousands of connections. In place of a single deep network for image recognition, this will enable an evolutionary hypernetwork to be set up, grouping together different patterns, with the possibility of integrating unknown patterns such as precarious or unfinished constructions, anarchic land use, flooding, etc.
Contribution / added value to the affiliated project :
In a first phase, the PhD research program will provide the project with a real-time monitoring, prediction and decision support tool for waste detection, localization and characterization. Precarious or unfinished constructions, anarchic land uses and the like could be addressed at the end of the PhD program.
Thesis supervisor Prof. Jean Marie DEMBELE, ACE MITIC, Senegal
Thesis co-supervisor Prof. Eugène EZIN, ACE SMIA, Benin
Other contributors to thesis supervision : Prof. Christophe CAMBIER, Sorbonne-Université - IRD