The recent wave of technological advances in machine learning and artificial intelligence has led to widespread applications and public awareness. At the same time, the rapid growth of high-speed wireless network services presents an opportunity for future distributed learning involving a vast number of smart IoT devices. This project targets several technical challenges posed by the limited reliability of wireless connections and computational constraints of the edge nodes in distributed learning systems. Overcoming these challenges is vital to the plethora of computation, communication, and coordination tasks required by distributed machine learning at the network edge. Centered on developing innovative edge learning algorithms over wireless MAC channels under the constraints of computing, power, and bandwidth, this project can significantly impact wireless edge learning in a variety of IoT applications, ranging from transportation, safety, and agriculture, to energy efficiency, e-health, and smart infrastructure. The broader impact of this research will also come through many educational opportunities by providing opportunities in STEM to K-12, women, and underrepresented minority students.
This collaborative project will develop an innovative network architecture for distributed learning over wireless multi-access channels. Specifically, the PIs will take a principled approach to develop an integrated wireless edge learning framework, using both gradient-based methods and also very recent advances in gradient-free, zero-order optimization, while taking into account the constraints in computing, power and bandwidth therein, in a holistic manner. The developed methods will be also extended to the setting of distributed online learning and reinforcement learning under wireless MAC. The PIs will focus on optimizing communication-efficient gradient sparsification based local updates that are communicated within the wireless network under bandwidth constraints; and each sender intelligently carries out transmission power allocation based on learning gradient and channel conditions. One important objective is to develop a novel learning-based framework for efficient wireless channel estimation and update to enable effective power control and learning. The project will devise edge learning algorithms that are robust against wireless channel uncertainty. The team of PIs shall comprehensively investigate the impact of the wireless bandwidth and power constraint on both the accuracy and convergence speed of edge learning algorithms.
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.
