How it started
I started my Ph.D. on the first of June 2017. The Ph.D. is under cifre convention with the participation of Ensta-Paris, U2IS Laboratory, and Thales Theresis. My supervisor from Ensta-Paris is David Filliat and my advisor from Thales was Jean-Francois Goudou and is now Andrei Stoian. The initial subject was “Incremental deep learning of detection and classification in a robotics setting”. The goal was to work on deep learning algorithms that can discover new objects or new instances of objects and learn to classify and detect them. The final goal was to apply it to a robot to explore a new environment.
Then
During my first year, I mostly worked on state representation learning (SRL) and generative models.
My idea was to use generative models as an active memory that can be sampled to create souvenirs for another learning model. My SRL work was a continuum of what I did before in my internship at ENSTA-Paris. The goal is to learn in an unsupervised fashion high dimension representation to help later RL algorithm, for example, to learn faster and transfer faster. Maybe there was some opportunity to mix SRL, generative models and continual learning later for some innovation …
At the beginning of my second year, I started using my work on generative models to test it on continual learning and trying to find some innovation for continual learning classification. After that, we can say that my thesis subject drifted to “Using Generative models for continual learning”.
Finally
My thesis will normally finish the 31 may, so we will see what happen until then.
Check out my scholar or HAL for more info on my articles. You may also want to see my github repositories.