The motivation of this blog, developped by my friend Corentin CARTEAU and myself, is to build intuition about the behaviour of state of the art optimization algorithms in the face of particular loss shapes. Our interactive simulator includes first order (Gradient Descent-like methods) and second order (Newton-like) methods on a wide range of test functions (Scale inhomogeneity, Local Minima, Non-Stationary functions...), as well as a Backtracking-Armijo Linesearch method. Click on the Figure to initiate an optimization procedure according to the specified parameters.
13 Apr 2020
The motivation of this blog post is to focus on the main rationale behind the use of nonparametric models in machine learning, and to specifically study one example of such model, the Dirichlet Process.
09 Sep 2019
In this post, we deal with datasets consisting of a social graph where each individual gives its top preferred other individuals in the network. The problematic is then to probabilistically model such a network in order to predict the behaviour of unknown individuals or infer network structures such as latent overlapping communities. We also discussed more refined situations where one can add covariates on individuals.
08 Sep 2019
This post present the report of my mini project of the course "Networks" at the university of Oxford. It deals with the apparent lack of performance of community detection of classical spectral methods and present the operators introduced in the litterature to overcome this problem.
07 Sep 2019
