Abstract
The rapid development of deep learning enables significant breakthroughs for intelligent edge‐terminal devices. However, neural network training for edge computing is currently overly dependent on cloud service platforms, resulting in low adaptivity for fast‐changing real‐world environments. The training energy efficiency is also strictly constrained by the traditional Von‐Neumann architecture with separate memory and processing units. To improve the adaptability and energy efficiency of edge‐terminal devices, a fully parallel online neural network training scheme based on electrochemical random‐access memory (ECRAM) arrays is proposed and validated. By exploiting the intrinsic linearity and nonlinearity functionalities of ECRAMs brought by varying numbers and amplitudes of programming pulses, a physical implementation of in situ multiplication using pulse‐based training is achieved, realizing fully parallel in situ computation and storage of outer product between two vectors. It can not only greatly accelerate the computation of weight gradients in neural network training but also significantly reduce the time complexity, latency, and energy overheads associated with data handling compared to traditional hardware implementations for this task. The ECRAM‐based online training system reduces the energy overhead of the training process by 30× when compared to the same training process executed on traditional computing hardware.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call