Samuel Prevost

A serverless-based app that collects public data off Facebook and Instagram and analyzes those bands, posts, images, and texts, to extract high-level features. Those features could include what the image contains, the angle at which the picture was taken, the sentiment and subjectivity of the post's text, and much more. The most innovative approach to this project was the implicit nature of those features, meaning the system wasn't trained on any specific set of labels and, therefore, was quite robust to handle any industry's posts.

In 2019, I worked on a joint R&D project with Sorbonne University and my current employer. The goal was to apply AI techniques to the field of art history to extract insights into how 19th-century painters worked. I directed a research team of junior data scientists, focusing on text analysis of painters' letters and style transfer technologies on their paintings.

We presented our work on style transfer at the NeurIPS | 2019 convention, and the project is now pending approval for a grant from the European Research Council. This project would revolutionize the field of art history by bringing new analysis techniques, never before used in the domain, and it could push the state of the art in style transfer and deep learning explainability.

A small Python project that essentially performs an inverse discrete Fourier transform on a given image in order to embed it into a sound in such a way that it can later be recovered from that sound's spectrogram. I wrote the first version of this project in 2017 to challenge my understanding of the Fourier transform, and it went viral when an influential AI author noticed and talked about it.

A C++ application that uses OpenCV to learn and recognize faces and associate them to names through a SQLite database. I wrote the application as part of an essay on biometrics, in which I went into detail on local binary pattern histograms (LBPH) and Haar Cascades (Viola-Jones face detection algorithm). This project was originally written in Python and then rewritten in C++ to run faster and on less powerful embedded systems.

A C++ implementation of Newton's method to minimize a multivariate objective function. Unlike stochastic gradient descent (SGD), the Newton method converges quadratically. This project also implements its own minimalist matrix library and inversion algorithm.

A C++ implementation of the Simplex algorithm, which optimizes any linear system of equations according to a linear objective function. The Simplex was implemented using the Eigen library, which is ubiquitous in C++. This implementation can detect common pitfalls such as unbounded problems.