FMSys Workshop 2025 Keynote

Measurements → Meaning → Action: Architecting Foundation Models for Cyber-Physical Intelligence

Speaker: Mani Srivastava (UCLA)

Abstract: Machine learning (ML) methods have profoundly transformed Cyber-Physical Systems and Internet-of-Things (CPS-IoT), shifting from reliance on first-principles mechanistic models and simple statistical methods toward sophisticated models learned directly from sensory observations. This transition fundamentally reshapes the perception-cognition-communication-action loops that characterize CPS-IoT. However, initial ML methodologies, primarily reliant on supervised learning with labeled data to construct task-specific models, face critical limitations. These challenges include difficulty scaling to heterogeneous sensor modalities, varied deployment scenarios, evolving application tasks, and dynamic operational conditions inherent to real-world CPS-IoT environments.

The remarkable successes of task-agnostic foundation models (FMs), particularly multimodal large language models (LLMs), in domains like natural language processing, computer vision, and speech recognition, have sparked considerable interest in leveraging these flexible models as generalizable building blocks within CPS-IoT analytics pipelines. Such foundation models promise significant reductions in costly and labor-intensive task-specific engineering. Nevertheless, a notable gap remains between current FM and LLM capabilities and the stringent performance, adaptability, and efficiency requirements inherent to CPS-IoT applications.

This talk explores this capability gap, drawing from insights and empirical results from our recent research. It identifies key requirements and design principles—captured by the progression from sensor measurements through meaningful interpretation to effective actions—that CPS-IoT domain-specific FMs and LLMs must meet to serve effectively as foundational tools for the next generation of cyber-physical intelligence.

Biography: Mani Srivastava is a Distinguished Professor, Mukund Padmanabhan Term Chair, and Vice Chair of Computer Engineering at UCLA's ECE Department with a joint appointment in the CS Department. He enjoys working on learning-enabled, resource-optimized, and trustworthy human-cyber-physical and IoT systems in the context of systems and applications for mobile health, smart environments, national security, and sustainability. His full-stack research examines challenges across the edge-cloud continuum in application mechanisms, architecture abstractions, algorithms, and platform technologies. He is a Fellow of the ACM and the IEEE. Srivastava also holds a concurrent appointment with Amazon AWS as an Amazon Scholar. The contents of this talk are based on his research at UCLA and unrelated to his work at Amazon.