Semester of Graduation
Spring 2026
Degree Type
Dissertation/Thesis
Degree Name
Neuromorphic Computing for Edge AI
Department
College of Computing and Software Engineering
Committee Chair/First Advisor
Liang Zhao
Second Advisor
Bobin Deng
Third Advisor
Shaoen Wu
Fourth Advisor
Yan Fang
Abstract
This dissertation investigates how spiking neural networks (SNNs) can improve federated edge intelligence by advancing three interconnected goals: communication efficiency, adversarial robustness, and continual adaptation. As edge computing deployments expand across Internet of Things (IoT), sensing, and privacy-sensitive applications, conventional federated learning approaches built around artificial neural networks (ANNs) face growing limitations in power consumption, bandwidth demand, and resilience to real-world uncertainty. SNNs offer an alternative computational paradigm based on event-driven, sparse, and temporally structured processing that is naturally suited to constrained edge environments. However, their behavior in practical federated settings remains insufficiently understood.
To address this gap, this dissertation develops a unified research agenda around federated neuromorphic intelligence. First, it studies noisy communication in federated learning and shows that SNNs are significantly more robust than equivalent ANN models when model updates are corrupted in transmission. Building on this finding, it proposes communication-efficient federated algorithms based on Top-$\kappa$ sparsification and dynamic-$\kappa$ reduction, demonstrating that SNN-based federated learning can maintain strong accuracy while reducing transmitted parameters to as low as approximately $6\%$ of the original model size. Second, it examines Byzantine robustness in federated SNNs and presents a statistical characterization explaining when SNNs are naturally resilient and when they are vulnerable, especially under structured adversaries such as MinMax. Motivated by this analysis, it introduces Top-$\kappa$ Vector-wise Trimming (TVT), a lightweight mechanism that jointly improves Byzantine robustness and reduces communication cost. Third, it broadens the robustness analysis by benchmarking surrogate-gradient-based federated SNNs against multiple Byzantine attacks and robust aggregation rules, showing that robustness is attack-dependent and that surrogate gradient choice materially affects recoverability. Fourth, it extends federated SNN research beyond static tasks by introducing a neuromorphic federated continual learning framework and demonstrating that SNNs can reduce catastrophic forgetting and improve final accuracy in no-replay federated continual learning settings. Finally, it addresses realistic deployment heterogeneity through a hybrid ANN--SNN federated learning framework, showing that collaborative training across conventional and neuromorphic clients can improve accuracy, accelerate convergence, reduce spike activity, and provide preliminary evidence of temporal robustness.
Taken together, the dissertation argues that the distinctive statistical and temporal properties of SNNs are not merely implementation details, but strategic advantages for distributed edge AI. The findings establish that neuromorphic models can be leveraged to make federated learning more communication-efficient, more adversarially robust, and more adaptable to evolving data and heterogeneous systems. This work contributes both practical algorithms and a broader design perspective for next-generation federated intelligence at the edge.