My current research interests fall within the very broad field of machine learning safety that aims at developing reliable, safe and trustworthy machine learning methods that can be deployed in the wild, wild world!

My PhD thesis is focused on uncertainty quantification in deep neural networks for agricultural computer vision. Particularly, I’m currently studying and working on conformal prediction, that allows us to transform any black box predictor into a set predictor with formal guarantess on the inclusion of the true value at a given confidence level (the idea is similar to that of confidence intervals – with some important differences).

My other research interests include domain adaptation, generalisation, and robustness of deep learning methods (which are all very tightly connected fields), as well as theories of learning (statistical learning, PAC learning, etc.), the history and philosophy of statistics and probability. On the applicative side, what motivates me the most are applications of machine learning in agriculture, the environment and broadly speaking anything related to earth and environmental sciences.

Publications

Peer-Reviewed Journals

Melki, P., Bombrun, L., Millet, E., Diallo, B., Elchaoui Elghor, H. & Da Costa, J.-P. (2022). Exploratory Analysis on Pixelwise Image Segmentation Metrics with an Application in Proximal Sensing. Remote Sensing.

Conferences & Workshops

Melki, P., Bombrun, L., Diallo, B., Dias, J. & Da Costa, J.-P. (2024). The Penalized Inverse Probability Measure for Conformal Classification. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. Seattle, WA, USA.
Melki, P., Bombrun, L., Diallo, B., Dias, J. & Da Costa, J.-P. (2023). Group-Conditional Conformal Prediction via Quantile Regression Calibration for Crop and Weed Classification . Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pp. 614-623. Paris, France.
Melki, P., Bombrun, L., Millet, E., Diallo, B., Elchaoui Elghor, H. & Da Costa, J.-P. (2023). A Study of Pixelwise Segmentation Metrics using Clustering of Variables and Self-Organizing Maps. Acta Horticulturae: XXXI International Horticultural Congress.