Panel Discussion:
Edge AI and Distributed Intelligence: Architectural Innov

Panelists: Soo-Kyun Kim, Jinsung Kim, Sangoh Park, Byoungjin Seok

Chair: Yanghoon Kim

The proliferation of Internet of Things devices, autonomous systems, and real-time applications has driven a fundamental shift from centralized cloud computing toward edge and distributed intelligence architectures. Edge AI processes data locally on end devices—smartphones, autonomous vehicles, industrial sensors, smart cameras—reducing latency, conserving bandwidth, enhancing privacy, and enabling operation in network-constrained or disconnected environments. Distributed intelligence coordinates multiple edge devices and edge servers to collaboratively execute complex tasks exceeding individual device capabilities. These architectural transformations promise to unlock transformative applications: autonomous vehicles that must make split-second decisions without cloud connectivity, augmented reality systems requiring sub-ten-millisecond response latency, privacy-preserving smart home devices that process sensitive data locally, and industrial automation systems operating in network-isolated factory environments.

However, edge AI deployment confronts substantial technical challenges. Resource constraints on edge devices—limited computational power, memory, energy budget—restrict deployable model complexity. Model compression techniques including quantization, pruning, knowledge distillation, and neural architecture search aim to reduce model size and computational requirements while preserving accuracy, yet trade-offs remain significant. Heterogeneity across edge devices complicates deployment—diverse processors, operating systems, and hardware accelerators require platform-specific optimizations. Distributed coordination challenges arise when multiple edge devices must collaborate—how to partition computation, synchronize model updates, and aggregate results while managing intermittent connectivity and Byzantine failures?

Security vulnerabilities multiply at the edge. Physical access to edge devices enables adversaries to extract models, manipulate inputs, or tamper with hardware. Federated learning, despite privacy benefits, faces poisoning attacks where malicious participants contribute corrupted model updates. Adversarial examples optimized for edge-deployed models threaten safety-critical applications like autonomous driving. Secure communication and trusted execution environments provide partial mitigation but introduce overhead and complexity.

This panel brings together experts in edge computing, machine learning systems, and distributed architectures to examine recent advances and persistent challenges in edge AI and distributed intelligence. Panelists will discuss:

Model Optimization for Edge Deployment: What compression techniques achieve optimal accuracy-efficiency trade-offs? How do quantization-aware training, structured pruning, and neural architecture search compare? What role will specialized AI accelerators (NPUs, TPUs, neuromorphic processors) play in edge AI evolution?

Distributed Learning Frameworks: How do federated learning, split learning, and peer-to-peer learning compare for different edge deployment scenarios? What optimization algorithms effectively train models across heterogeneous, resource-constrained devices? How can systems maintain model quality despite non-IID data distributions across devices?

Edge-Cloud Collaboration: What workload partitioning strategies optimally divide computation between edge and cloud? How can systems dynamically adjust partitioning based on network conditions, device resources, and application requirements? What coordination protocols minimize communication overhead?

Security and Privacy: How can edge AI systems resist model extraction, adversarial examples, and poisoning attacks? What trusted execution environment technologies are mature enough for production edge deployment? How do privacy-preserving techniques like differential privacy and secure aggregation perform under edge resource constraints?

Real-World Deployments: What lessons emerge from production edge AI systems in autonomous vehicles, smart cities, industrial IoT, and healthcare? What unforeseen challenges emerged during deployment? What success stories demonstrate edge AI potential?

Future Directions: How will 5G/6G networks with edge computing capabilities transform edge AI architectures? What role will emerging technologies like neuromorphic computing and photonic AI accelerators play? How can standardization efforts facilitate interoperable edge AI ecosystems?

Through case study presentations and interactive discussion, this panel will synthesize state-of-the-art research findings with practical deployment insights, providing attendees with comprehensive understanding of edge AI opportunities and challenges as distributed intelligence becomes increasingly prevalent across application domains.