ML

BCAM’s Machine Learning research tackles several methodological axes that jointly showcase the group’s distinctive and comprehensive approach to Machine Learning.

▹ Robustness and Learning Under Uncertainty

We develop minimax and probabilistic ML frameworks that provide rigorous theoretical guarantees for reliable predictive performance. This ensures systems remain robust even when deployed with data that is noisy, scarce, imbalanced, evolving, or subject to adversarial manipulation.

Key methodological directions include:

• Minimax Learning and Robust Classification: Developing optimal strategies for robustness against worst-case data scenarios and adversarial manipulation.
• Performance Guarantees Under Distribution Shift: Designing methods with theoretical resilience against covariate shift and concept drift.
• Fairness and Stability: Ensuring fairness is maintained even under unstable or uncertain data-generating processes.
• Uncertainty-Aware Robustness: Applying probabilistic learning techniques to handle challenges like data missingness and model uncertainty.

Examples: ML methods for evolving tasks with guarantees, fairness stability analysis, formal characterization of adversarial attacks on models and explanations.
 

▹ Adaptive and Online Learning in Dynamic Environments and Time-Series

Time-series Machine Learning is a central area of expertise at BCAM. We design algorithms capable of efficiently adapting and updating themselves as conditions change, which is critical for dynamic domains like energy forecasting, physiological process modeling, industrial monitoring, and complex sensor networks.

Research efforts and contributions include:

• Bayesian Time-Series Analysis: modeling temporal dependencies, quantifying uncertainty in forecasts, and performing reliable inference in dynamic environments.
• Time-series anomaly detection: via self-supervised representation learning and generative modeling
• Online and streaming time-series: prediction and classification, as well as classification, imputation, and forecasting in multivariate sensor data
• Elastic-, distance- and kernel-based time-series modeling and learning
• Learning and forecasting with hybrid (statistical and machine learning) dynamical system
• Adaptive strategies for concept drift
• Early and incremental classification for real-time systems

Examples: surveys on time-series anomaly detection, online DTW, adaptive load forecasting, selective imputation for sensor data, early and elastic classification, streaming novelty detection, time-series meta-learning.
 

▹ Probabilistic Machine Learning and Uncertainty Quantification

Many real-world problems require not only a description or prediction, but also a rigorous understanding of the confidence or uncertainty associated with such outcome. Our research in this area is rooted in probability and statistical theory, providing a solid mathematical bedrock for advanced ML methodologies with calibrated uncertainty quantification.

Our methodological contributions are structured around:

Foundational Generative Modeling:

• Developing Probabilistic Generative Models (e.g., Normalizing Flows, Variational Autoencoders) designed to understand the underlying data-generating process, enabling descriptive, predictive, and prescriptive frameworks.
• Designing Graphical and Structured Probabilistic Models (including Bayesian networks) for complex data dependencies.

Principled Treatment of Data Challenges:

• Handling Missingness and Data-Drift: Providing a principled treatment of missingness (with or without imputation) and data-distribution drifts using probabilistic methods.
• Weak and Noisy Supervision: Advanced learning under weak- or noisy supervision, addressing scenarios like learning from label proportions, positive-unlabeled data, or biased/unreliable annotators.

Algorithm Development and Scope:

• Efficient Inference: Creating efficient algorithms for learning and inference in high-dimensional and complex models: Variational Inference, Markov Chain Monte Carlo, Sequential Monte Carlo, etc.
• Broad Data Scope: Methodologies for modeling diverse data types, including continuous, discrete, ranking, and time-series data.

Examples: Bayesian classifiers from label proportions, Gaussian Processes, Normalizing Flows, probabilistic treatment models, semi-supervised multi-class learning.
 

▹ Trustworthy and Explainable Machine Learning

Our research contributes methodological advancements that make ML systems interpretable, reliable, and accountable for high-stakes decision-makers. This ensures trust and transparency from model development through deployment:

Key contributions focus on:

• Fairness, Stability, and Transparency: Developing methods and theoretical guarantees that ensure algorithmic fairness and model stability, particularly when data distributions are shifting or uncertain.
• Explainable AI for Complex Data: Creating explainable, statistical and Machine Learning models specifically tailored for complex data.
• Trustworthy ML: Rigorously analyzing the vulnerabilities of both machine learning models and their explanations (e.g., adversarial manipulation of explanations), for robust system design.

Examples: Explainability in time series, adversarial manipulation of explanations, fairness under shifting distributions and its stability analysis, feature attribution methods.
 

▹ Federated Learning of Classifiers and Probabilistic Models

We design distributed and decentralized algorithms for training classifiers and probabilistic models over heterogeneous, privacy-sensitive networks.

• Federated and decentralized learning algorithms
• Learning under non-IID, drifting, or adversarial data conditions
• Privacy and robustness in distributed systems

Examples: decentralized learning of probabilistic classifiers, learning over time-varying communication networks, robust aggregation under adversarial or unreliable nodes.
 

▹ Sequential Decision-Making Under Uncertainty

We research the mathematical grounds for designing intelligent agents that offer calibrated measures of predictive and prescriptive uncertainty, which is essential for systems operating in high-stakes environments, where actions must be taken based on imperfect, noisy and evolving information.

Our methodological focus revolves around reinforcement learning (RL) and control theory:

• Multi-Armed Bandits (MAB): Developing theory and algorithms for the optimal trade-off between exploration and exploitation in complex, real-world settings.
• Decision Making in Non-Stationary Environments: Extending classical decision-making under uncertainty for dynamic and non-stationary environments, to robustly adjust to drifting reward distributions and evolving dynamics.
• Multi-Objective and Contextual Optimization: Devising algorithms that can optimize for sequential and potentially conflicting goals (e.g., maximizing returns while minimizing uncertainty or risk) and integrate rich contextual information to enable informed sequential choices under uncertainty.

Examples: Online resource allocation and scheduling, adaptive clinical trial design, dynamic pricing algorithms.
 

▹ Unified Vision

By integrating robust theoretical learning, probabilistic modeling, time-series methodology, trustworthy ML, federated learning, and sequential decision-making, the BCAM Machine Learning research line provides a coherent and comprehensive scientific foundation for the safe and effective development and deployment of AI.

Our objective extends beyond advancing ML theory: we ensure these mathematically grounded methodological advances translate into reliable, interpretable, and high-impact applications across strategic domains, including energy, health, industry, and the environment.