ERI: Back Propagation-Free Machine Learning for Split Neural Networks in Distributed Edge Systems

This NSF ERI project aims to make collaborative machine learning more practical for real-world edge systems where data are distributed across devices, and networks often differ in speed, reliability, and computing capability. Today, many distributed learning methods require each device to train a full neural network or to exchange large amount of information during training, which can be costly for edge devices such as wearables, mobile devices, and other resource-limited platforms. The project will develop a new class of split learning methods that avoid the heavy communication required by standard back propagation (BP), while still allowing devices and servers to train models together without sharing raw data. This will be achieved by replacing repeated gradient exchanges with lightweight scalar updates that are better suited for heterogeneous multi-edge environments. The intellectual merit of the project includes establishing the theoretical foundations, algorithmic designs, and evaluation methods needed for back propagation-free split learning in distributed systems. The broader impacts of the project include expanding access to advanced machine learning for organizations and communities with limited computing and networking resources, supporting education and workforce development through research-integrated training, and releasing open-source tools and benchmarks that can benefit the broader research and education communities. Technically, this project studies how to train split neural networks when the client side does not receive full gradients propagated back from the server. In conventional split learning, a neural network is divided between edge devices and a server so that devices compute early layers and the server computes later layers. While this reduces device-side computation compared with federated learning, it normally requires two-way communication of intermediate activations and gradients at every training step. This project will develop back propagation-free split learning methods that use forward-only scalar loss signals to update the client-side model. This research project has three integrated goals: (1) to derive convergence guarantees and communication-accuracy tradeoffs for these methods; (2) to design adaptive algorithms that remain robust under heterogeneous device capabilities and changing bandwidth conditions; and (3) to build reproducible benchmarks for comparing BP-based and BP-free approaches across representative models and datasets. The expected contribution is a new foundation for scalable, communication-efficient, and privacy-aware collaborative learning in distributed edge systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2552997 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Chao Huang | Institution: Montclair State University, MONTCLAIR, NJ | Award Amount: $199,494 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2552997 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2552997.html