February 9 – 13, 2026 — Machine Learning Winter School 2026 
Seville, Spain 
San Diego, USA 
Hangzhou, China 
Sharm El Sheikh, Egypt 
Rotterdam, Netherlands 
Vienna, Austria 
United Kingdom 
Vancouver, Canada 
Yokohama, Japan MBZUAI Nexus Speaker Series
Upcoming AI Talks
Past AI Talks
Reinforcement Learning in Health-related Sequential Decision Problems: From Dynamic Treatment Regimes to Mobile Health
Hosted by: Prof. Marcos Matabuena
February 13, 2026
1:00pm
MBZUAI - Classroom 4
Bibhas Chakraborty
Associate Professor and Deputy Director in Biomedical Data Science at Duke-NUS Medical School
Hosted by: Prof. Marcos Matabuena
Computational Biology
Join Stream
Abstract
Reinforcement Learning in Health-related Sequential Decision Problems: From Dynamic Treatment Regimes to Mobile Health
In recent years, Reinforcement Learning (RL) has gained a prominent position in addressing health-related sequential decision-making problems. In this talk, we will discuss two such sequential decision-making problems: (1) dynamic treatment regimes (DTRs), i.e., clinical decision rules for adapting the type, dosage and timing of treatment according to an individual patient’s characteristics and evolving health status; and (2) just-in-time adaptive interventions (JITAIs) in mobile app-based behavioral nudges in population health. Specifically, we will illustrate the similarities and differences between these two types of RL problems (e.g., offline vs. online RL), common algorithms used in these two settings (e.g., Q-learning vs. Thomson sampling), and real-life case studies.
In recent years, Reinforcement Learning (RL) has gained a prominent position in addressing health-related sequential decision-making problems. In this talk, we will discuss two such sequential decision-making problems: (1) dynamic treatment regimes (DTRs), i.e., clinical decision rules for adapting the type, dosage and timing of treatment according to an individual patient’s characteristics and evolving health status; and (2) just-in-time adaptive interventions (JITAIs) in mobile app-based behavioral nudges in population health. Specifically, we will illustrate the similarities and differences between these two types of RL problems (e.g., offline vs. online RL), common algorithms used in these two settings (e.g., Q-learning vs. Thomson sampling), and real-life case studies.
MLMC: Visualizing Multi-Label Classification. A Tool for Intuitively Evaluating and Comparing Classifiers at Global, Label and Instance Levels
Hosted by: Prof. Laura Koesten
February 13, 2026
11:00am
MBZUAI - Classroom 1
Torsten Möller
Professor of Computer Science at University of Vienna
Hosted by: Prof. Laura Koesten
Human-Computer Interaction
Join Stream
Abstract
MLMC: Visualizing Multi-Label Classification. A Tool for Intuitively Evaluating and Comparing Classifiers at Global, Label and Instance Levels
Machine learning classifiers are increasingly applied to complex tasks such as audio tagging, image labeling, and text classification -- many of which require multi-label classification. Traditional evaluation tools, often limited to single metrics such as accuracy, fall short of providing insight into classifier behavior across multiple labels. To address this, we present MLMC, an interactive visualization tool for evaluating and comparing multi-label classifiers. Based on expert interviews, MLMC supports analysis at instance-, label-, and classifier-level views, offering a scalable, more interpretable alternative. We demonstrate its use across three different domains and describe its core algorithms and user interface. Two pilot studies (N=$6$ each) provided insight into MLMC's usability and showed improved task accuracy, consistency, and user confidence compared to confusion matrices. Results highlight MLMC's potential as a practical tool for intuitive evaluation of multi-label classifiers, with implications for a broad range of machine learning applications. Our approach is using the Design Study Methodology, which is rooted in Human-Centered Design.
Machine learning classifiers are increasingly applied to complex tasks such as audio tagging, image labeling, and text classification -- many of which require multi-label classification. Traditional evaluation tools, often limited to single metrics such as accuracy, fall short of providing insight into classifier behavior across multiple labels. To address this, we present MLMC, an interactive visualization tool for evaluating and comparing multi-label classifiers. Based on expert interviews, MLMC supports analysis at instance-, label-, and classifier-level views, offering a scalable, more interpretable alternative. We demonstrate its use across three different domains and describe its core algorithms and user interface. Two pilot studies (N=$6$ each) provided insight into MLMC's usability and showed improved task accuracy, consistency, and user confidence compared to confusion matrices. Results highlight MLMC's potential as a practical tool for intuitive evaluation of multi-label classifiers, with implications for a broad range of machine learning applications. Our approach is using the Design Study Methodology, which is rooted in Human-Centered Design.
SDE Matching: Simulation-Free Learning of Stochastic Dynamics
Hosted by: Prof. Eric Moulines
February 16, 2026
11:00am
MBZUAI - Classroom 5
Christian Andersson Naesseth
Assistant Professor of Machine Learning at University of Amsterdam
Hosted by: Prof. Eric Moulines
Machine Learning
Join Stream
Abstract
SDE Matching: Simulation-Free Learning of Stochastic Dynamics
"Stochastic differential equations (SDEs) provide a flexible framework for modeling time series, dynamical systems, and sequential data. However, learning SDEs from data typically relies on adjoint sensitivity methods, which require repeated simulation, time discretization, and backpropagation through approximate SDE solvers, leading to significant computational overhead and limited scalability. We introduce SDE Matching, a simulation- and discretization-free approach for learning stochastic dynamics directly from data. Building on recent advances in score matching and flow matching for generative modeling, we extend these ideas to the dynamical setting, enabling direct learning of SDE drift and diffusion terms without numerical simulation. SDE Matching replaces solver-based training with a regression-like objective defined on transformed data samples, eliminating the need for backpropagation through stochastic trajectories. Empirically, SDE Matching achieves accuracy comparable to adjoint sensitivity-based methods while substantially reducing computational cost, offering a scalable alternative for learning stochastic dynamical systems. We demonstrate these results across a range of synthetic and real-world dynamical modeling tasks."
"Stochastic differential equations (SDEs) provide a flexible framework for modeling time series, dynamical systems, and sequential data. However, learning SDEs from data typically relies on adjoint sensitivity methods, which require repeated simulation, time discretization, and backpropagation through approximate SDE solvers, leading to significant computational overhead and limited scalability. We introduce SDE Matching, a simulation- and discretization-free approach for learning stochastic dynamics directly from data. Building on recent advances in score matching and flow matching for generative modeling, we extend these ideas to the dynamical setting, enabling direct learning of SDE drift and diffusion terms without numerical simulation. SDE Matching replaces solver-based training with a regression-like objective defined on transformed data samples, eliminating the need for backpropagation through stochastic trajectories. Empirically, SDE Matching achieves accuracy comparable to adjoint sensitivity-based methods while substantially reducing computational cost, offering a scalable alternative for learning stochastic dynamical systems. We demonstrate these results across a range of synthetic and real-world dynamical modeling tasks."
Information Design for the Information Age
Hosted by: Prof. Eric Moulines
February 18, 2026
10:30am
MBZUAI - Lecture Hall 2
Safwan Hossain
Ph.D candidate in Computer Science at Harvard University
Hosted by: Prof. Eric Moulines
Machine Learning
Join Stream
Abstract
Information Design for the Information Age
Information design is a seminal concept in economics wherein a party with information advantage can strategically reveal this to influence the actions of a rational decision-maker. This talk centers on my efforts to bridge this model to emerging computational and machine learning paradigms. While the classic model assumes that only the quantitative structure of information matters, behavioral economics and psychology emphasize that the framing of information also plays a key role. My recent work formalizes a language-based notion of framing for information design and combines analytical methods to design information structures with LLMs to optimize the language/framing. I explore, both theoretically and empirically, when this LLM-augmented approach is tractable. I will also discuss a second work that uses information design as a light-weight approach to content moderation on social media. Doing so requires a new framework where the information advantage originates from a machine learning model and the interaction is dynamic with long-term intervention effects. I will conclude by connecting these threads to my broader research agenda on strategic decision-making in multi-agent systems.
Information design is a seminal concept in economics wherein a party with information advantage can strategically reveal this to influence the actions of a rational decision-maker. This talk centers on my efforts to bridge this model to emerging computational and machine learning paradigms. While the classic model assumes that only the quantitative structure of information matters, behavioral economics and psychology emphasize that the framing of information also plays a key role. My recent work formalizes a language-based notion of framing for information design and combines analytical methods to design information structures with LLMs to optimize the language/framing. I explore, both theoretically and empirically, when this LLM-augmented approach is tractable. I will also discuss a second work that uses information design as a light-weight approach to content moderation on social media. Doing so requires a new framework where the information advantage originates from a machine learning model and the interaction is dynamic with long-term intervention effects. I will conclude by connecting these threads to my broader research agenda on strategic decision-making in multi-agent systems.
Stochastic optimal control approach to generative modelling and Schrödinger potential estimation
Hosted by: Prof. Eric Moulines
February 19, 2026
11:00am
MBZUAI - Classroom 4
Alexey Naumov
Professor and director of AI institute at HSE University in Moscow
Hosted by: Prof. Eric Moulines
Machine Learning
Join Stream
Abstract
Stochastic optimal control approach to generative modelling and Schrödinger potential estimation
Stochastic optimal control problem with a final constraint provides a natural way to construct a Schrödinger bridge between two distributions, making it well‑suited for generative modelling. In this problem, the optimal control can be expressed through the Schrödinger potential, which depends on the target distribution — typically unknown in practice. We address the problem of estimating this potential from finite samples. Focusing on estimators that minimize the empirical Kullback Leibler (KL) divergence, we study their generalization abilities. Despite the loss function’s unusual structure, we show that it exhibits favourable geometric properties under mild assumptions that hold for a broad class of target distributions. We derive non‑asymptotic, high‑probability upper bounds for the potential estimation accuracy, measured in terms of excess KL‑risk. In the second part of the talk , we show that the Schrödinger system could be rewritten in terms of a single positive transformed potential that satisfies a nonlinear fixed-point equation and estimate this potential by empirical risk minimization over a function class. The talk is based on the joint work with D. Belomestny, N. Puchkin and D. Suchkov.
Stochastic optimal control problem with a final constraint provides a natural way to construct a Schrödinger bridge between two distributions, making it well‑suited for generative modelling. In this problem, the optimal control can be expressed through the Schrödinger potential, which depends on the target distribution — typically unknown in practice. We address the problem of estimating this potential from finite samples. Focusing on estimators that minimize the empirical Kullback Leibler (KL) divergence, we study their generalization abilities. Despite the loss function’s unusual structure, we show that it exhibits favourable geometric properties under mild assumptions that hold for a broad class of target distributions. We derive non‑asymptotic, high‑probability upper bounds for the potential estimation accuracy, measured in terms of excess KL‑risk. In the second part of the talk , we show that the Schrödinger system could be rewritten in terms of a single positive transformed potential that satisfies a nonlinear fixed-point equation and estimate this potential by empirical risk minimization over a function class. The talk is based on the joint work with D. Belomestny, N. Puchkin and D. Suchkov.
High-Dimensional Multi-Agent Robot Learning
Hosted by: Prof. Yoshihiko Nakamura
February 23, 2026
11:00am
MBZUAI - Classroom 5
Guillaume Adrien Sartoretti
Assistant Professor in Mechanical Engineering at National University of Singapore (NUS)
Hosted by: Prof. Yoshihiko Nakamura
Robotics
Join Stream
Abstract
High-Dimensional Multi-Agent Robot Learning
As robotic systems grow more capable and ubiquitous, their increasing scale and complexity necessitate a shift toward robust, scalable controllers and automated synthesis methods. My group has approached this challenge by turning to distributed (multi-agent) reinforcement learning (MARL) approaches, with an emphasis on understanding and eliciting emergent coordination/cooperation in multi-robot systems and articulated robots (where agents are individual joints). There, our focus lies in improving information representations and neural architectures, as well as devising learning techniques that can help them explore their high-dimensional joint policy space, to identify and reinforce high-quality policies that naturally fit together towards team-level cooperation. In this talk, I will discuss the three main areas my group has been investigating: imitation learning, modularized/hierarchical neural structures, and learning scaffolding. I will describe these techniques within a wide variety of robotic applications, such as multi-agent pathfinding, autonomous exploration/search, traffic signal control, collaborative manipulation, and legged loco-manipulation. Finally, I will also briefly touch on some of our ongoing and future work. Throughout this journey, my goal will be to highlight the key challenges surrounding learning representation, policy space exploration, and scalability/robustness of learned policies, and outline some of the open avenues for research in this exciting area of robotics.
As robotic systems grow more capable and ubiquitous, their increasing scale and complexity necessitate a shift toward robust, scalable controllers and automated synthesis methods. My group has approached this challenge by turning to distributed (multi-agent) reinforcement learning (MARL) approaches, with an emphasis on understanding and eliciting emergent coordination/cooperation in multi-robot systems and articulated robots (where agents are individual joints). There, our focus lies in improving information representations and neural architectures, as well as devising learning techniques that can help them explore their high-dimensional joint policy space, to identify and reinforce high-quality policies that naturally fit together towards team-level cooperation. In this talk, I will discuss the three main areas my group has been investigating: imitation learning, modularized/hierarchical neural structures, and learning scaffolding. I will describe these techniques within a wide variety of robotic applications, such as multi-agent pathfinding, autonomous exploration/search, traffic signal control, collaborative manipulation, and legged loco-manipulation. Finally, I will also briefly touch on some of our ongoing and future work. Throughout this journey, my goal will be to highlight the key challenges surrounding learning representation, policy space exploration, and scalability/robustness of learned policies, and outline some of the open avenues for research in this exciting area of robotics.
What We Can Know: Statistical Learning From Complex Functional Data
Hosted by: Hongyuan Cao
March 19, 2026
11:00am
MBZUAI - Classroom 5
Shichang Zhang
Ian McKeague
Hosted by: Hongyuan Cao
Statistics and Data Science
Join Stream
Abstract
What We Can Know: Statistical Learning From Complex Functional Data
This talk introduces a novel nonparametric inference framework for functional data having sample paths of bounded variation, with applications in a variety of complex statistical settings. The main application will be to wearable device data collected in a Columbia-based study of an experimental therapy for mitochondrial disease, a group of disorders that affect the body's ability to produce energy. Specifically, we provide the first clinical application of a novel, bias-adjusted outcome measure of acceleration across a range of subjects' activities to assess nucleoside therapy for thymidine kinase 2 deficiency, an ultra-rare autosomal recessive mitochondrial disease.
This talk introduces a novel nonparametric inference framework for functional data having sample paths of bounded variation, with applications in a variety of complex statistical settings. The main application will be to wearable device data collected in a Columbia-based study of an experimental therapy for mitochondrial disease, a group of disorders that affect the body's ability to produce energy. Specifically, we provide the first clinical application of a novel, bias-adjusted outcome measure of acceleration across a range of subjects' activities to assess nucleoside therapy for thymidine kinase 2 deficiency, an ultra-rare autosomal recessive mitochondrial disease.
