UC San Diego

The rapid growth of Machine Learning (ML) model sizes poses challenges for mobile applications, especially when compared to the slower pace of mobile hardware development. Although cloud-based ML lightens this load by moving computations to Data Center (DC) servers, it struggles with limited bandwidth. On the other hand, Mobile Edge Computing (MEC)-based ML offers faster response times but can’t always handle intense computations. To find a balance, split ML is introduced, which distributes ML tasks across various computing platforms.

The study delves into the inherent challenges of split ML, proposing innovative solutions: (1) HiveMind, a split ML system optimized for cellular networks; (2) NeuroMessenger, a mechanism that uses ML’s inherent error tolerance to reduce data transmission delays; and (3) X-Array, an innovative radio architecture that meets split ML’s high bandwidth needs. Together, these contributions seek to enhance the efficiency and feasibility of ML in the mobile computing landscape.