Adrien Le Coz

Characterization of a Reliability Domain for Image Classifiers

The goal of my thesis was to explore methods for defining a reliability domain that would clarify the conditions under which a model is trustworthy. Three aspects have been considered:

Qualitative: Generating synthetic extreme examples helps illustrate the limits of a classifier and better understand what causes it to fail.

Quantitative: Selective classification allows the model to abstain when uncertain, and calibration improves uncertainty quantification.

Semantic: Multimodal models can provide textual descriptions of images likely to lead to incorrect or, conversely, to correct predictions.

Confidence Calibration of Classifiers with Many Classes

Adrien Le Coz, Stéphane Herbin, Faouzi Adjed

NeurIPS 2024

We transform the problem of calibrating a multiclass classifier into calibrating a single surrogate binary classifier and show that it significantly improves existing calibration methods.

Efficient Exploration of Image Classifier Failures with Bayesian Optimization and Text-to-Image Models

Adrien Le Coz, Houssem Ouertatani, Stéphane Herbin, Faouzi Adjed

Generative Models for Computer Vision – CVPR 2024 Workshop

We propose an image classifier benchmarking method as an iterative process that alternates image generation, classifier evaluation, and attribute selection. This method efficiently explores the attributes that ultimately lead to poor behavior.

Explaining an image classifier with a generative model conditioned by uncertainty

Adrien Le Coz, Stéphane Herbin, Faouzi Adjed

Uncertainty meets Explainability – ECML-PKDD 2023 Workshop

We propose to condition a generative model by a given image classifier uncertainty in order to analyze and explain its behavior.

Leveraging generative models to characterize the failure conditions of image classifiers

Adrien Le Coz, Stéphane Herbin, Faouzi Adjed

Artificial Intelligence Safety – IJCAI-ECAI 2022 Workshop

We characterize the failure conditions of image classifiers and generate corner cases by expressing failure conditions as directions in the latent space of StyleGAN2.

Towards optimal district heating temperature control in China with deep reinforcement learning

Adrien Le Coz, Tahar Nabil, Francois Courtot

Tackling Climate Change with Machine Learning – NeurIPS 2020 Workshop

We propose a deep reinforcement learning agent to optimize the temperature control of district heating systems in China. The agent has been tested on a real-world system.

Machine learning based design of a supercritical CO2 concentrating solar power plant

Tahar Nabil, Yann Le Moullec, Adrien Le Coz

European supercritical CO2 Conference 2019

In an approach similar in principle to current LLMs, we represent thermodynamic power cycles as sequences of strings and train a recurrent neural network to predict the next character of the sequence. The network is then used to autoregressively sample new cycles.