Open-World & Robust Machine Learning
Out-of-distribution detection, open-set recognition, uncertainty and robustness for safe deployment of ML models.
Deployed models constantly face inputs unlike their training data. This line develops methods that let ML models know what they do not know: detecting out-of-distribution (OOD) inputs, recognizing unknown classes in data streams, quantifying uncertainty, and balancing performance, efficiency and robustness.
The topic grew out of two earlier lines of my work: concept drift in data streams, where the data distribution changes over time, and the adversarial robustness of time series and image classifiers. Since 2021 it has become a line of its own, and it is the subject of Aitor Martínez-Seras’s doctoral thesis on OOD detection in open-world machine learning (2024). The work covers:
- definitions and taxonomies of what is unknown to a model;
- OOD detection for conventional, spiking and object-detection networks, and in reinforcement learning;
- open-set recognition in data streams;
- uncertainty quantification with conformal prediction and topological data analysis;
- domain generalization;
- trade-offs between accuracy, efficiency and robustness when models are compressed for deployment.
Main contributions
- Foundations: a unifying review of how to manage the unknown in machine learning, covering definitions, related areas, recent advances and prospects (Barcina-Blanco et al., 2024).
- OOD detection: a detector for spiking neural networks with built-in explainability (Martínez-Seras et al., 2023), an analysis of the inherent robustness of one-stage object detectors against OOD data (Martínez-Seras et al., 2026), the limits of neural collapse for OOD detection (Terres-Escudero et al., 2026), post-hoc explanations as OOD detectors (Martinez et al., 2022), open-set object discovery in autonomous driving (Martínez-Seras et al., 2023), and detection and adaptation strategies for OOD observations in reinforcement learning (Martínez-Seras et al., 2024).
- Open-set recognition over data streams (Barcina-Blanco et al., 2024).
- Uncertainty quantification: neural activations and uncertainty in object detection transformers through topological data analysis (Del Ser et al., 2024), confidence of traffic forecasting models (Laña et al., 2024), the connection between concept drift and uncertainty (Lobo et al., 2023), and conformal methods for motion prediction and annotation disagreement (Tamuly et al., 2026; Villate-Castillo et al., 2025).
- Robustness and generalization: adversarial attacks and defences for time series and image classifiers (Oregi et al., 2018; Oregi et al., 2020), domain generalization (Gao et al., 2025; Xiong et al., 2024), and evolutionary multi-objective compression and pruning that balance accuracy, efficiency and robustness (IJCNN 2024 Best Paper Award) (Del Ser et al., 2024; Poyatos et al., 2023).
Publications in this area (19)
All my publications related to this research line, in reverse chronological order, including those mentioned above.
2026
2025
2024
- 2024 International Joint Conference on Neural Networks (IJCNN)Balancing Performance, Efficiency and Robustness in Open-World Machine Learning via Evolutionary Multi-Objective Model CompressionIn 2024 International Joint Conference on Neural Networks (IJCNN), 2024
2023
- IEEE International Conference on Intelligent Transportation Systems (ITSC)