Internships

Assessing modularity of deep generative models (January 2018 to June 2018)

Under the supervision of Michel Besserve at the Empirical Inference department at the Max Planck Institute for Intelligent Systems (Tuebingen, Germany)

5 month and a half long exploratory internship into the study of deep generative model with respect to the Independence of Cause and Mechanism postulate from causality theory.
The general goal was to quantify the modularity in generative networks by considering the effect of interventions at a given layer on the next layer, with approaches ranging from a spectral quantification of the effect of shifts in the Fourier domain to direct interventions on activation maps of convolutional layers.
My end-of-internship report can be accessed here and a preliminary workshop paper on the effect of direct interventions was accepted at the Workshop on Theoretical Foundations and Applications of Deep Generative Models at ICML 2018.

Detecting and highlighting domain shifts from output distribution in a multiclass setting (October 2017 to December 2017)

Under the supervision of Christoph Lampert in his CVML group at the Institute for Science and Technology (IST) Austria (Klosterneuberg, Austria)

12 week long exploratory internship into the detection of cases where a classifier is used on data that differs too much from what it was trained for.
The main idea was to detect shifts relevant to the classifier (as opposed to random shifts in the data that have nothing to do with what the classifier uses for classification) by considering the outputs of a classifier. Relevant shifts were assumed to translate into a change of the underlying distribution of classifier softmax outputs and statistical tests were used to detect those changes in distribution.
My end-of-internship report can be accessed here and a preliminary paper on the simplest method developed was received for an oral presentation at the German Conference on Pattern Recognition (GCPR 2018, anciently DAGM).

Deep Learning and latent representations of peptidic sequences (May 2016 to July 2016)

Under the supervision of François Coste in team Dyliss at IRISA (Rennes, France)

8 week long exploratory internship into the possibilities of applying Deep Learning techniques to Bioinformatics
Revolved around the use of auto-encoding recurrent structures to acquire representations of amino acids in the context of a particular protein superfamily and the associated latent representation space
My end-of-internship report can be accessed here.