José Lucas De Melo Costa
PhD candidate in machine learning, LISN, CentraleSupélec, Université Paris-Saclay
costa@lisn.frGoogle Scholargithub.com/jose-meloORCID 0009-0007-5372-9101Massy, France
Updated October 2026 · PDF version
Research
I study the training dynamics of self-supervised learning, mainly joint-embedding predictive architectures (JEPAs), using linearized and spectral models to predict when representations collapse (NeurIPS 2026, ICLR 2025). A second line attaches distribution-free uncertainty to tabular models through conformal prediction, with anomaly and fraud detection as the main application (ESANN 2026; best paper, MultiTab at MICCAI 2026). Before the PhD I worked on spectral and graph methods for sensor and medical time series (OMAE 2021, 2024).
Seven conference papers, six as first author, including NeurIPS 2026 and ICLR 2025; three workshop and forum papers.
Education
2024 – 2027
PhD in Computer Science, CentraleSupélec, Université Paris-Saclay
LISN, GALaC team. Supervisors: Bich-Liên Doan and Fabrice Popineau. Thesis topic: self-supervised learning and anomaly detection for tabular data.
2021 – 2023
Master's in Engineering, major in artificial intelligence, CentraleSupélec
2018 – 2023
B.Sc. in Computer Engineering, Escola Politécnica, University of São Paulo
Research experience
2024 – present
PhD researcher, LUSIS–CentraleSupélec chair, LISN and LUSIS
- JEPAs for tabular data, without data augmentation (ICLR 2025).
- A linearized, per-mode spectral analysis of JEPA training whose stability ratio predicts the collapse boundary across 800+ configurations (NeurIPS 2026).
- Named convergence collapse in fixed-target anomaly detectors and a kernel-anchored regularizer that mitigates it (CIKM 2026).
- Conformal uncertainty for tabular foundation models: a 112-dataset benchmark (ESANN 2026) and single-pass calibration for cross-modal screening (Best Paper, MultiTab at MICCAI 2026).
- Self-supervised fraud detection on LUSIS card-transaction data, with a public-benchmark version presented at the Institut Louis Bachelier forum (2026).
Jun – Oct 2024
Research engineer, LISN, CentraleSupélec
Survey of anomaly-detection models and self-supervised baselines for tabular data.
Apr – Sep 2023
Research intern, efficient vision transformers, CEA, Saclay
Halved vision-transformer size at equal accuracy with CKA-guided quantization.
Jun – Sep 2022
Research intern, time-series anomaly detection, Université Paris-Cité
Graph-based representations of multivariate medical time series for anomaly detection.
Apr – Dec 2020
Research intern, offshore structural monitoring, Petrobras and University of São Paulo
Mooring-line failure detection from platform motion, by natural-frequency estimation and later spectral graph networks; two OMAE papers.
Publications
Conference papers
† equal contribution[1]
J. L. De Melo Costa, S. W. Ahn, F. Popineau, A. Rimmel, B.-L. Doan. Drive vs. Decay: On the Training Dynamics of Joint-Embedding Predictive Architectures. NeurIPS 2026, main track.
[2]
J. L. De Melo Costa, F. Popineau, A. Rimmel, B.-L. Doan. Mitigating Convergence Collapse in Fixed-Target Anomaly Detectors via Kernel-Anchored Locality Regularization. CIKM 2026, full research track.
[3]
S. W. Ahn, A. Leite, J. L. De Melo Costa, F. Popineau, B.-L. Doan, A. Rimmel. Knowledge-Informed Local Causal Discovery of Optimal Adjustment Sets. CIKM 2026, full research track.
[4]
H. Thimonier†, J. L. De Melo Costa†, F. Popineau, A. Rimmel, B.-L. Doan. T-JEPA: Augmentation-Free Self-Supervised Learning for Tabular Data. ICLR 2025.
[5]
J. L. De Melo Costa, F. Popineau, A. Rimmel, B.-L. Doan. High Performance, Low Reliability: Uncertainty Benchmarking for Tabular Foundation Models. ESANN 2026.
[6]
J. L. De Melo Costa, F. Kurike Matsumoto, C. F. D. Netto, M. R. de Barros, A. do N. Queiroz Filho, E. S. Gomi, E. A. Tannuri, A. H. Reali Costa. Spectral Graph-Based Networks for Mooring Line Failure Detection on FPSO. ASME OMAE 2024.
[7]
J. L. De Melo Costa, A. N. Queiroz Filho, I. H. F. Santos, R. A. Barreira, A. H. Reali Costa, E. S. Gomi, E. A. Tannuri. FPSO Mooring Line Integrity Supervising System Based on Motion Data and Natural Frequency Estimation. ASME OMAE 2021.
Workshop and forum papers
[8]
J. L. De Melo Costa, S. W. Ahn, F. Popineau, A. Rimmel, B.-L. Doan. Single-Pass Conformal Cross-Modal Anomaly Screening with Tabular Foundation Models. MultiTab workshop at MICCAI 2026, oral, best paper award.
Springer LNCS 17263project pagecode
[9]
J. L. De Melo Costa, C. Moineau, T. Allenet, I. Kucher. Centered Kernel Alignment for Efficient Vision Transformer Quantization. AccML workshop at HiPEAC 2024, oral.
[10]
J. L. De Melo Costa, F. Popineau, A. Rimmel, B.-L. Doan. Leveraging Self-Supervised Learning for Fraud Detection in Tabular Data. Financial Risks International Forum, Institut Louis Bachelier, 2026.
Under review and in revision
[11]
S. W. Ahn, J. L. De Melo Costa, A. Leite, R. Minato, F. Popineau, A. Rimmel, B.-L. Doan. A Diagnostic Radius for Knowledge-Induced Causal Effect Estimation. Submitted to ICLR 2027.
[12]
J. L. De Melo Costa, F. Popineau, A. Rimmel, B.-L. Doan. TabSets: Benchmarking Commitment, Confidence and Label-Free Sets of Tabular Foundation Models. Neurocomputing, in revision.
Invited extension of [5].
Awards
2026
Best Paper Award, MultiTab workshop at MICCAI 2026, Strasbourg
2023
Ranked 1st of 48 in the entering class, B.Sc. Computer Engineering, University of São Paulo
2022
AUCANI Academic Merit Scholarship, university-wide, University of São Paulo
Teaching
2026 – 2027
Teaching assistant, CentraleSupélec. Introduction to Python Programming (M1).
2025 – 2027
Teaching assistant, CentraleSupélec. Algorithms and Complexity.
2024 – 2025
Teaching assistant, CentraleSupélec. Introduction to Python Programming; Algorithms.
2019
Teaching assistant, University of São Paulo
Algorithms and Data Structures; Systems Programming.
Service and supervision
2025 – 2026
Reviewer for Neurocomputing and the NeurIPS 2026 Trust-AI-Eval workshop. Supervisor of a CentraleSupélec 3rd-year industrial project for the LUSIS chair (three students, 2025–26).
Skills and languages
Methods
Self-supervised learning, JEPAs, training dynamics, spectral analysis, conformal prediction, anomaly detection, tabular foundation models, kernel methods, graph neural networks.
Tools
Python, PyTorch, scikit-learn, C/C++, LaTeX, Slurm.
Languages
Portuguese (native), English (C1, Cambridge Linguaskill), French (C1, TCF).