Grant: Discovery Grants Program - Individual
Université de Montréal
$23,900.00
Advancing Machine Learning with Causal Models
FY 2023-24
Summary
Machine learning (ML) is a key ingredient for intelligent agents that help us with science and decision making. However, ML faces hurdles that limit its wider use: (i) ML systems struggle to generalize out-of-distribution (OOD), to unseen tasks and domains; (ii) ML systems learn correlations but science and decision making require causal inference -- inference about the effects of interventions. The field of causality, with its formalism of causal models, provides a theoretical framework to address these shortcomings and benefits from ML too: instead of carefully measuring variables of interest and defining causal models, with ML, we can learn infer quantities from rich sources of data. The long-term objective of the proposed research program is to facilitate scientific progress and decision making by combining the theoretical framework of causality with flexible approaches from modern ML. To this end, we focus on three short-term objectives: Objective 1: Representation learning for causal inference To learn variables of interest from data with ML requires representation learning, the task of mapping observations to unobserved latent variables. However, representation learning is not guaranteed to recover the true underlying latents. In this objective, we will develop new theory and methods for representation learning that suffice for valid downstream causal inference. Objective 2: Causal models with latent structure ML methods offer the promise of learning causal models from experimental and non-experimental (i.e., observational) data. However, there are combinatorially many models to consider, and state-of-the-art approaches are still limited to discovering causal models over tens of variables, well below the scale needed for science. The goal of this objective is to scale up the discovery of causal models. The key insight will be to exploit structure in the data that drastically reduces the number of models to search over. Objective 3: Learning causal mechanisms from multiple environments Recent methods combine ML and data from changing (observational) data distributions, called environments, to learn causal mechanisms. While these results are promising for causality, they suffer from inefficient data-use and unrealistic assumptions. In this objective, we will develop a new efficient and flexible framework for learning causal mechanisms from multiple environments. The framework rests on a novel insight: environment changes do not happen by chance; environments evolve according to causal processes. We will show how causal models over environments lead to efficient algorithms for learning causal mechanisms from data. The proposed research program will impact both fundamental ML research as well as the development of intelligent agents for scientific discovery and decision making.The program will also train HQP to be experts in causality and ML, providing them foundational tools to thrive in both industry and academic roles.
Program
Discovery Grants Program - Individual
Recipient
Location
Quebec, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data
As published
Application id
RGPIN-2023-04869
Name nom
Sridhar, Dhanya
Department département
Département d'informatique et de recherche opérationnelle (DIRO)
Organization id
63
Institution établissement
Université de Montréal
Province en
Québec
Province fr
Québec
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2023
Competition year année de concours
2023
Award amount
23900
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1507
Committee name en
Computer Science
Committee name fr
Sciences informatiques
Area of application code
1207
Area of application group en
Advancement of knowledge
Area of application group fr
Avancement des connaissances
Area of application en
Information, computer and communication technologies
Area of application fr
Technologie de l'information, des ordinateurs et des réseaux de communications
Research subject code
9999
Research subject group en
Not available
Research subject group fr
Not available
Research subject en
Non disponible
Research subject fr
Non disponible
Application title
Advancing Machine Learning with Causal Models
Keyword
Causality; Machine learning; Probabilistic models; Representation learning
Application summary
Machine learning (ML) is a key ingredient for intelligent agents that help us with science and decision making. However, ML faces hurdles that limit its wider use: (i) ML systems struggle to generalize out-of-distribution (OOD), to unseen tasks and domains; (ii) ML systems learn correlations but science and decision making require causal inference -- inference about the effects of interventions. The field of causality, with its formalism of causal models, provides a theoretical framework to address these shortcomings and benefits from ML too: instead of carefully measuring variables of interest and defining causal models, with ML, we can learn infer quantities from rich sources of data. The long-term objective of the proposed research program is to facilitate scientific progress and decision making by combining the theoretical framework of causality with flexible approaches from modern ML. To this end, we focus on three short-term objectives: Objective 1: Representation learning for causal inference To learn variables of interest from data with ML requires representation learning, the task of mapping observations to unobserved latent variables. However, representation learning is not guaranteed to recover the true underlying latents. In this objective, we will develop new theory and methods for representation learning that suffice for valid downstream causal inference. Objective 2: Causal models with latent structure ML methods offer the promise of learning causal models from experimental and non-experimental (i.e., observational) data. However, there are combinatorially many models to consider, and state-of-the-art approaches are still limited to discovering causal models over tens of variables, well below the scale needed for science. The goal of this objective is to scale up the discovery of causal models. The key insight will be to exploit structure in the data that drastically reduces the number of models to search over. Objective 3: Learning causal mechanisms from multiple environments Recent methods combine ML and data from changing (observational) data distributions, called environments, to learn causal mechanisms. While these results are promising for causality, they suffer from inefficient data-use and unrealistic assumptions. In this objective, we will develop a new efficient and flexible framework for learning causal mechanisms from multiple environments. The framework rests on a novel insight: environment changes do not happen by chance; environments evolve according to causal processes. We will show how causal models over environments lead to efficient algorithms for learning causal mechanisms from data. The proposed research program will impact both fundamental ML research as well as the development of intelligent agents for scientific discovery and decision making.The program will also train HQP to be experts in causality and ML, providing them foundational tools to thrive in both industry and academic roles.