Jose Costa
PhD Student · LISN, CentraleSupelec, Paris-Saclay University
I work on self-supervised learning for tabular data, anomaly detection, and robust machine learning systems. My recent research focuses on JEPA training dynamics and fraud detection in real-world financial data.
News
Publications
Self-Supervised Learning & Joint-Embedding Architectures
Drive vs. Decay: On the Training Dynamics of Joint-Embedding Predictive Architectures
Costa, J.L.M., Ahn, S.W., Popineau, F., Rimmel, A., Doan, B.-L.
NeurIPS 2026 (poster).
Why joint-embedding predictive architectures collapse in the first epochs of training, and how identity-biased predictors keep their representations from decaying.
Leveraging Self-Supervised Learning for Fraud Detection in Tabular Data
Costa, J.L.M., Popineau, F., Rimmel, A., Doan, B.-L.
19th FINANCIAL RISKS INTERNATIONAL FORUM, Institut Louis Bachelier.
Pretraining on unlabelled transactions improves credit-card fraud detection in the regime that makes it hard: almost no labels, and fewer than one fraud in a thousand.
T-JEPA: Augmentation-Free Self-Supervised Learning for Tabular Data
Thimonier, H.*, Costa, J.L.M.*, Popineau, F., Rimmel, A., Doan, B.-L.
ICLR 2025. *Joint first authors.
Self-supervised learning for tables without hand-crafted augmentations, by predicting the latent representation of masked subsets of features.
Uncertainty Quantification & Anomaly Detection
Mitigating Convergence Collapse in Fixed-Target Anomaly Detectors via Kernel-Anchored Locality Regularization
Costa, J.L.M., Popineau, F., Rimmel, A., Doan, B.-L.
Oral presentation at CIKM 2026, Rome.
Anomaly detectors that regress towards a fixed target get worse as they converge; their kernel analogue is a closed-form, training-free detector, and anchoring the network to it keeps the anomaly signal alive under prolonged training.
Single-Pass Conformal Cross-Modal Anomaly Screening with Tabular Foundation Models
Costa, J.L.M., Ahn, S.W., Popineau, F., Rimmel, A., Doan, B.-L.
Oral presentation at MICCAI 2026 @ MultiTab Workshop, Strasbourg.
Masking a tabular foundation model's attention diagonal recovers every leave-one-out calibration score in a single forward pass, making cross-modal anomaly screening far better calibrated at no extra training cost.
High Performance, Low Reliability: Uncertainty Benchmarking for Tabular Foundation Models
Costa, J.L.M., Popineau, F., Rimmel, A., Doan, B.-L.
ESANN 2026, pp. 115-120.
Tabular foundation models predict accurately but quantify their uncertainty badly — a benchmark showing that strong performance does not imply trustworthy confidence.
Causal Discovery & Effect Estimation
Knowledge-Informed Local Causal Discovery of Optimal Adjustment Sets
Ahn, S.W., Leite, A., Costa, J.L.M., Popineau, F., Doan, B.-L., Rimmel, A.
CIKM 2026, Rome.
Injecting expert edge constraints directly into local causal discovery, and propagating them with Meek's rules, recovers optimal adjustment sets that observational data alone leaves unidentifiable.
Applied Machine Learning for Engineering Systems
Spectral Graph-Based Networks for Mooring Line Failure Detection on FPSO
Costa, J.L.M. et al.
ASME 2024 43rd International Conference on Ocean, Offshore & Arctic Engineering (OMAE), Singapore.
A spatial-temporal graph neural network that spots mooring-line failures on floating oil platforms from the platform's motion alone.
FPSO Mooring Line Integrity Supervising System Based on Motion Data and Natural Frequency Estimation
Costa, J.L.M., Queiroz Filho, A.N., Santos, I.H.F., Barreira, R.A., Costa, A.H.R., Gomi, E.S., & Tannuri, E.A.
ASME 2021 40th International Conference on Ocean, Offshore and Arctic Engineering (OMAE).
Detects mooring-line failure on offshore platforms by tracking shifts in the platform's natural oscillation frequencies.
Efficient & Interpretable Models
Centered Kernel Alignment for Efficient Vision Transformer Quantization
Costa, J.L.M., Moineau, C., Allenet, T., & Kucher, I.
Oral presentation at 6th AccML Workshop, HiPEAC 2024, Munich.
Aligning a network's internal representations lets a vision transformer be quantized to half its size while keeping its accuracy.
Learning and Understanding Strategies on Zero-Sum Games
Costa, J.L.M., Poli, J., Nguyen, M.-B., Ducatez, A., Bouscary, M., & Boucher, B.
Oral presentation at Junior Multidisciplinary Congress, Universite Paris-Saclay, 2022.
What machine-learning agents actually learn when they play zero-sum games, and how far those strategies can be explained.