Quentin Delfosse
Machine Learning Group, Computer Science Department, TU Darmstadt. Hochschulstrasse 1, Room S1|03 077, 64289 Darmstadt, Germany
quentindelfosse.me
quentin (dot) delfosse (at) cs (dot) tu-darmstadt (dot) de
Meetings by appointment.

Mission. We show that current reinforcement learning agents are not able to adapt to task simplification. I focus on creating RL agents that leverage human inductive biases to learn meaningful policies, that can be used on simpler environments. Here is an outlook of my PhD research.

Timeline.
2020 - now: Ph.D. student at the Machine Learning Lab, CS Department, TU Darmstadt, Germany
2017 - 2019: M.Sc. MoSIG in Computer Science at UGA Grenoble (Double Diploma), France
2016 - 2019: Dipl. Ing. in Computer Science at ENSIMAG Grenoble, France
2014 - 2016: Higher School Preparatory Class at Lycée Buffon Paris, France


Projects.

Rational Activation Functions

Rationals as learnable activation functions, trainable in an end-to-end fashion using backpropagation.

Neural Plasticity for Deep RL

Allowing learning structure to embed high plasticity (with rationals) is crucial for dynamic RL environments.

Object Detection with Time Consistency

Using space-time consistency to better detect and understand objects.

Comic Style Transfer

Transforming real images into comic style ones using Generative Adversarial Networks.

GANCA

Combining Generative Adversarial Networks with Neural Cellular Automata.

Supervised Theses.
2021 Siyao Chen, Comic Generation with Rational Generation Model, M.Sc. Thesis, co-supervision Arseny Skryagin
2021 Maximilian Otte, Creating Emojis with Generative Adversarial Neural Cellular Automata, M.Sc. Thesis, co-supervision Johannes Czech
2021 Julius Zimmermann, Evaluation and Showcasing of Rational Functions, M.Sc. Thesis, co-supervision Patrick Schramowski
2020 Patrick Vimr, Optimizing the analysis of Rational Networks applied to Deep Reinforcement Learning-Algorithms, B.Sc. Thesis

Supervised Courses and Projects.
WS 2020/21 Quentin Delfosse, Patrick Schramowski, Improving the Rational Activation Functions library, Bachelor Praktikum