Tutorial 3: Dynamic Architectural Adaptations for Resource-Efficient Generative AI in IoT Systems

Speakers: Divya Saxena and Jiannong Cao

Duration: 3hr

Abstract: Generative AI has emerged as a transformative force in modern computing, enabling advancements in autonomous systems, smart infrastructure, and personalized healthcare. Its ability to generate synthetic data, adapt to new domains with minimal supervision, and model complex patterns makes it a powerful tool for enhancing sensor data, enabling predictive modeling, and supporting adaptive learning in IoT systems. However, traditional generative AI models, while effective, often require high computational resources, large memory footprints, and abundant training data, making them impractical for resource-constrained IoT environments characterized by low power budgets, limited storage, and data scarcity. This tutorial introduces resource-efficient generative AI techniques that go beyond existing efficient AI frameworks by focusing on dynamic architectural adaptation. While existing methods provide static compression or once-off model pruning, dynamic architectural adaptation methods allow on-the-fly optimization of network architecture during training, enabling flexibility and adaptability to data constraints without requiring extensive retraining. Additionally, they optimize memory usage and reduce computational overhead, enabling generative AI models to be scalable, adaptive, and deployable in resource-limited environments. Participants will gain practical insights into designing, adapting, and deploying generative AI models for tasks such as synthetic data generation and domain adaptation in IoT scenarios. The tutorial is practically focused, showcasing the design, optimization, and deployment of generative AI models for real-world applications, including multi-modal MRI translation and food quality detection—both critical tasks in resource-constrained IoT systems such as remote healthcare monitoring and supply chain automation. We will conclude with insights into emerging trends such as diffusion-based generative models and continual adaptation, exploring how these approaches can enhance flexibility and long-term adaptability in IoT applications. Attendees will learn to develop scalable and resource-efficient generative AI solutions ready for next-generation IoT deployments.

Speaker Bios

Dr. Divya Saxena (Senior Member, IEEE) is a Research Assistant Professor at The Hong Kong Polytechnic University, specializing in generative AI, adaptive architectures, and resource-efficient AI models. She earned her Ph.D. in Computer Science from IIT Roorkee, India, and has over six years of post-PhD experience focused on generative modeling, domain adaptation, and data-efficient learning techniques. Her research primarily addresses resource constraints in IoT and embedded systems by developing dynamic architectural adaptations for scalable generative AI solutions. Dr. Saxena has published in top-tier conferences such as CVPR, AAAI, and WACV, showcasing advancements in synthetic data generation, image-to-image translation, and few-shot learning techniques.

Prof. Jiannong Cao (Fellow, IEEE) received the Ph.D. degree in computer science from Washington State University, Pullman, WA, USA, in 1990. He is currently the Otto Poon Charitable Foundation Professor of Data Science and the Chair Professor of Distributed and Mobile Computing with the Department of Computing, The Hong Kong Polytechnic University (PolyU), Hong Kong, where he is also the Dean of the Graduate School, the Director of the Research Institute for Artificial Intelligence of Things, and the Director of the Internet and Mobile Computing Laboratory. His research interests include distributed systems and blockchain, big data and machine learning, wireless sensing and networking, and mobile cloud and edge computing.

Tutorial outline

Part 1: Introduction and Motivation (20 minutes)

Overview of Generative AI and Applications

  • Brief history and evolution of generative AI (GANs, VAEs, Diffusion Models)
  • Key areas of impact: autonomous systems, healthcare, smart infrastructure

Challenges in IoT Systems

  • Resource constraints (power, storage, data)
  • Scalability and latency limitations

Motivational Use Cases

  • Why resource-efficient generative AI is crucial for next-generation IoT
  • Outline of tutorial scope and learning objectives

Part 2: Fundamentals of Generative AI (30 minutes)

Core Concepts

  •  GANs: architecture, training dynamics, and variants
  •  VAEs: probabilistic modeling and latent space representations
  •  Diffusion Models: overview and recent advances

IoT-Specific Considerations

  •  Data scarcity and streaming data
  •  Handling sensor noise and real-time data flows

Application Examples

  •  Synthetic data generation for sensor calibration
  •  Domain adaptation across heterogeneous IoT devices

Part 3: Dynamic Architectural Adaptations (40 minutes)

Scalable and Adaptive Generative AI

  •  Introduction to pruning-regrowth strategies
  •  Selective fine-tuning to minimize retraining costs

On-the-Fly Optimization

  •  Dealing with limited, noisy, or diverse IoT data
  •  Short Code Demo: Show a pruning-regrowth implementation snippet
  •  Interactive Q&A (5 minutes): Answer participant questions or discuss code strategies

Techniques in Practice

  •  Performance comparisons and trade-offs

Part 4: Real-World Applications (40 minutes)

Case Studies

  •  Multi-modal MRI Translation
  •  Food Quality Detection

Applications in Practice

  •  Code snippets and live demonstrations

Part 5: Emerging Trends and Future Directions (30 minutes)

Diffusion-Based Generative Models

  •  Potential for increased stability and fidelity
  •  Suitability for IoT constraints and data-limited scenarios

Continual Learning for IoT

  •  Avoiding catastrophic forgetting
  •  Adapting to evolving IoT data distributions

Part 6: Interactive Q&A and Discussion (20 minutes)

  • Open Q&A Session