Brief Bio
I do research on Machine Learning and Computational Neuroscience, with a focus on Gradient Stability. We found it is an effective framework to understand spiking neurons. I’m personally grateful to Wolfgang Maass and Alexandre Pouget for the opportunity to do an internship at their labs. We also found that gradient explosion was misunderstood for deep RNNs, since an additive exponential explosion was never correctly described. In that work we also propose pretraining to stability as an effective strategy to improve generalization ability of an unconstrained range of deep RNNs. We built the HoME dataset in collaboration with MILA, Aaron Courville and Hugo Larochelle, for multimodal embodied learning. We also designed U-BESD, a neural network to help the hearing impaired isolate the sound they are trying to attend to. I recently submitted my thesis at Université de Sherbrooke under the supervision of Jean Rouat, and I am defending on April 10th 2024.
Essentially I’ve used several variants of Deep Learning for my research, including recurrent, convolutional, transformers, neuromorphic architectures and state-space models. I’ve used matrix norms, eigenvalues and Random Matrix theory to understand the stability of these architectures. I’m currently refining my last work using Probability Theory concepts such as Chebyshev’s inequality, and Analytic Combinatorics tools such as the saddle point method, generating functions and coefficient extractors to understand the stability of deep RNNs.
I want to mathematically describe cognition and come up with mathematical theories that can predict architecture quality before training. Somehow, I would like to use it to solve at least one disease :)
Contact me
luca.celotti at usherbrooke.ca
luca.herrtti at gmail.com