Sridhar Mahadevan

Sridhar Mahadevan

Adobe Systems (United States)
2 previous affiliations
126Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Sridhar Mahadevan focuses on the mathematical foundations of reinforcement learning, causal inference, and representation learning. His early work explored robotic skill composition and temporal difference learning, establishing robust frameworks for sequential decision-making. Recently, his research has shifted toward applying category theory and topology to unify machine learning architectures and causal models.

Research Trajectory

Mahadevan's early research addressed the practical challenges of scaling reinforcement learning (RL) to physical systems. In 1991, he demonstrated how to overcome sparse rewards and state history encoding issues by integrating RL with a subsumption architecture in Scaling Reinforcement Learning to Robotics by Exploiting the Subsumption Architecture, enabling a physical robot to learn component behaviors for box-pushing tasks. He subsequently formalized foundational aspects of RL, including his work on Average Reward Reinforcement Learning: Foundations, Algorithms, and Empirical Results (1996). Extending decision-making models to parallel execution, he collaborated on Learning to Take Concurrent Actions (2002), which investigated semi-Markov Decision Processes for agents executing temporally extended actions simultaneously, contributing to the broader area of Concurrent Composition of Robotic Skills.

In the mid-2000s, his focus expanded to the automatic discovery of representations in Markov Decision Processes. He introduced Proto-value Functions: A Laplacian Framework for Learning Representation and Control in Markov Decision Processes (2007), a significant contribution to the research direction of Laplacian State Representations for Option Discovery. This work utilized graph Laplacians to build basis functions that capture the underlying topology of the state space.

During the 2010s, Mahadevan addressed core algorithmic stability and optimization challenges in RL. With collaborators, he developed Proximal Reinforcement Learning: A New Theory of Sequential Decision Making in Primal-Dual Spaces (2014). By leveraging Legendre transforms and mirror maps, this framework provided mathematically rigorous approaches to off-policy stability and sparse representation, advancing the area of Primal-Dual Reinforcement Learning via Convex Reformulation. He continued this trajectory with Regularized Off-Policy TD-Learning (2020), introducing the RO-TD algorithm to mitigate instability and dimensionality issues, further establishing methods in the direction of Stable Off-Policy Temporal Difference Learning.

Most recently, Mahadevan has undertaken a comprehensive effort to unify machine learning, generative AI, and causal inference using advanced mathematics, particularly category theory and topos theory. In 2024, he proposed GAIA: Categorical Foundations of Generative AI, a theoretical blueprint that models neural architectures as hierarchical, simplicial structures rather than sequential chains. This work anchors his contributions to the broader area of Categorical and Topological Foundations for Unified Machine Learning. He expanded on this structural approach in Topos Theory for Generative AI and LLMs (2025), presenting a compositional framework for next-generation architectures based on set-like spaces. Returning to reinforcement learning through this new mathematical lens, he authored Universal Reinforcement Learning in Coalgebras: Asynchronous Stochastic Computation via Conduction (2025), which reframes optimal value function search as a distributed coinduction process to unify MDPs, POMDPs, and PSRs.

Concurrently, Mahadevan has applied these categorical frameworks to reconceptualize causal inference, driving the research direction of Mathematical Foundations for Unified Causal Inference. In Universal Causal Inference in a Topos (2025), he generalized Structural Causal Models using topos theory, demonstrating that causal logic is fundamentally intuitionistic. He formalized this non-Boolean approach to interventions in Intuitionistic j-Do-Calculus in Topos Causal Models (2025), enabling stable causal inference across varying experimental regimes. To operationalize these theories, he introduced Decentralized Causal Discovery using Judo Calculus (2025), a framework that glues context-dependent local causal truths into global models for more scalable inference. Finally, bridging causality with modern language models, he developed DEMOCRITUS in Large Causal Models from Large Language Models (2025), a system that extracts and synthesizes causal claims from LLMs into navigable manifolds of cause-and-effect relationships.

Current Directions

Mahadevan's current research is heavily focused on rebuilding the mathematical foundations of artificial intelligence using category theory, topos theory, and universal coalgebras. Recent work investigates how hierarchical, simplicial structures can replace traditional sequential neural network layers to create more unified and compositional generative models.

A major theme in his latest papers is the generalization of causal inference. By moving from classical Boolean logic to intuitionistic logic within a topos, he is developing frameworks like $j$-do-calculus and Judo Calculus. These approaches aim to synthesize context-dependent, local causal mechanisms into coherent global models, enabling decentralized causal discovery and the construction of "Large Causal Models" mined from the latent knowledge of large language models. Additionally, he is exploring how universal coalgebras can provide a rigorous basis for asynchronous, distributed reinforcement learning across diverse environments.

Profile updated

See something wrong? Report a mistake