Highly Uniform Two‐Terminal Artificial Synapses Based on Polycrystalline Hf0.5Zr0.5O2 for Sparsified Back Propagation Networks. (15th July 2020)

• Record Type: Journal Article
• Authors: Lu, Yingming; Liu, Keqin; Yang, Jingjing; Zhang, Teng; Cheng, Caidie; Dang, Bingjie; Xu, Liying; Zhu, Jiadi; Huang, Qianqian; Huang, Ru; Yang, Yuchao
• Is Part Of: Advanced Electronic Materials. Volume 6:Number 8(2020)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2020-07-15
• Subjects: hafnium zirconium oxide -- in‐memory computing -- memristors -- online learning -- sparsified back propagation; Materials -- Electric properties -- Periodicals; Materials science -- Periodicals; Magnetic materials -- Periodicals; Electronic apparatus and appliances -- Periodicals; 537
• DOI: 10.1002/aelm.202000204
• Languages: English
• ISSNs: 2199-160X

Abstract

Abstract: Artificial neural networks with nonvolatile memristors as synapses represent an in‐memory computing paradigm, where the online training of networks with rapidly growing scale places high requirements on device uniformity and endurance as well as training time and energy. Here, a highly uniform two‐terminal Hf0.5 Zr0.5 O2 based artificial synapse is reported, which shows 4‐bit weight precision, cycle‐to‐cycle variation (σ/ µ ) of <4%, device‐to‐device variation (σ/ µ ) of <9%, retention of >10 4 s at 85 °C, and endurance of >10 6 . The improved uniformity can be attributed to the nonfilamentary resistive switching mechanism mediated by ion exchange at the interface. A holistic optimization using such reliable synapses and modified algorithm with sparsified back propagation can accelerate training by 166 times, decrease the energy consumption by 83 times, reduce the total updating rate by 138 times in multilayer perceptron, and the approach can be extended to convolutional neural networks as well. The synergistically optimized approach thus paves the way for the construction of memristor based systems capable of learning and interacting adaptively. Abstract : A two‐terminal polycrystalline Hf0.5 Zr0.5 O2 ‐ based synapse with excellent device‐to‐device and cycle‐to‐cycle uniformity is demonstrated experimentally via nonfilamentary resistive switching. Combination of highly uniform synapse with sparsified back propagation demonstrates a holistic solution to accelerate the training process, decrease energy/time consumption, and relax the requirement on device endurance, as shown in both multilayer perception and convolutional neural networks.