Yannick Le Cacheux

An app that I developed for Awen.ai. I was the lead data scientist on this project. The objective was to help creative professionals find ideas and iterate quickly on existing ones using generative AI. My work included modifying open-source image generation models, such as Stable Diffusion, and adding the ability to quickly integrate new concepts (objects, styles, etc.) within the model's knowledge base. I also integrated and fine-tuned large language models, such as Bloom and GPT-3.

A model that I developed for Saint Gobain. I was the lead data scientist for this project. The objective was to optimize the portfolio of sales representatives based on geographic criteria while balancing constraints (e.g., minimum and maximum size, revenue, etc.) per portfolio. While the problem formulated is NP-complete, a satisfying solution can be found using a greedy heuristic based on linear programming.

A machine learning model that I developed for L'Oréal. I was the lead scientist on this project. The model's objective was to rapidly estimate the 3D shape of a subject's face from one or several photos. It consists of an iterative approach based on statistical shape modeling. First, key points are detected in the photo(s), followed by the orientation, scale, and more. Then, the most likely principal components are iteratively estimated until convergence, which is typically reached in four steps. The model was integrated on iOS with Core ML.

A deep learning model that I developed for CEA. A zero-shot classifier aims to recognize images of classes for which no training instance is provided, based only on semantic information. For example, even without ever seeing a tiger, most humans should be able to recognize one if provided with the information that "a tiger is a big cat with orange fur and black stripes on its back."

Zero-shot classifiers can be useful when one does not have training data for all classes.

I was the lead scientist on this research project aiming to improve existing zero-shot classifiers. I, together with two other deep learning researchers, showed that most existing approaches lacked desirable theoretical properties. More specifically, the usual loss functions do not enable the model to capture certain intra-class and inter-class structures. We provided novel theoretical results and developed a new model capable of outperforming all previous existing models.

Our proposed approach was published and presented at the 2019 International Conference on Computer Vision in Seoul, Korea.

A predictive model that I developed for AXA. The main objective was to identify clients with several unlinked entries in internal databases, for instance, because they independently purchased several insurance products. This is essential from a client's knowledge point of view.

At first glance, this is neither very difficult nor exciting: if two people have the same name and address, they are the same person. Problem solved. Except homonyms exist, addresses change, typos are made, and databases can be messy. And with millions of clients, a brute-force pairwise comparison is not an option.

Hence the need for clever predictive models to efficiently and accurately identify duplicates. I was in charge of the data science team designing the models and assessing associated risks.

The regularly updated information provided by these models is now a central component in many internal information systems.

The project was presented at the 2016 Viva Technology show in Paris.