In the last decades, data and computing power has become ubiquitous. Consequently, neural networks have regained a lot of interest, though their inception dates back from the 50’s. They have indeed been proven to be ‘universal approximators’ meaning that under the limits of infinite data and neurons they can approximate any non-linear function arbitrarily well.
Today Deep Learning has revolutionized computer tasks such as image and speech recognition. In the fields of Identification and Control of Dynamical Systems, there is a growing interest for the use of Neural Networks in modelling and process control tasks. Indeed it has the potential to revolutionize complex system modelling in fields such as climate science, neuroscience or biology.
Although capable of reaching extraordinary performance, Deep Learning suffers from 3 drawbacks preventing its use in industrial applications: generalization, extrapolation and interpretation. In this thesis, we propose to address those challenges.
In this thesis, you will:
- Learn about Neural Networks, how they are built and how they ‘learn’.
- Study the current state of the art in dynamical modelling methods using deep learning.
- Study and propose different approaches to make a Neural Network interpretable and generalizable
- Implement state of the art algorithms in Python and Matlab for dynamical system modelling
- Propose and implement new methods in Python and Matlab to address the challenges of interpretability and generalization of Neural Networks applied to dynamical system modelling.
You will develop the following skills:
- scientific litterature study
- industrial practices in software development
- Deep Learning simulation
Want to know more ? Contact Antoine Marchal