The Characteristics of Binary Spike-Time-Dependent Plasticity in HfO2-Based RRAM and Applications for Pattern Recognition

Zheng Zhou, Chen Liu, Wensheng Shen, Zhen Dong, Zhe Chen, Peng Huang, Lifeng Liu, Xiaoyan Liu, Jinfeng KangNanoscale 2017 · DOI

TL;DR

Researchers have developed a novel neuromorphic computing system using HfO2-based RRAM devices that mimics the brain's learning mechanism (STDP) to efficiently recognize handwritten digits. This work demonstrates a promising pathway for creating more energy-efficient and powerful AI hardware by leveraging the unique properties of resistive switching memory.

What Problem Does This Paper Address?

Brain-inspired computing architectures hold immense promise for revolutionizing AI, particularly in areas like pattern recognition and big data processing, due to their potential for high efficiency and novel processing capabilities. A key component in these architectures is the artificial synapse, which needs to replicate the dynamic learning behavior of biological synapses. While memristive devices, specifically oxide-based RRAM, have emerged as strong candidates for these artificial synapses due to their non-volatility, scalability, and CMOS compatibility, existing RRAM-based Spike-Time-Dependent Plasticity (STDP) systems often require numerous training epochs for effective learning. This paper addresses the need for more efficient and faster learning in RRAM-based neuromorphic systems by proposing and demonstrating a binary STDP protocol that achieves excellent learning performance with fewer training iterations.

Methods

The core of this work lies in the development and experimental validation of a binary Spike-Time-Dependent Plasticity (STDP) protocol implemented using HfO2-based RRAM devices. STDP is a learning rule where the timing of pre- and post-synaptic spikes determines the change in synaptic strength. In this binary STDP protocol, the RRAM device's resistance state (representing synaptic weight) is modified based on the precise temporal overlap between incoming "pre-spikes" and outgoing "post-spikes."

The researchers fabricated an 8x16 RRAM crossbar array using a Pt/Al2O3/HfO2/TiN/Al structure. The fabrication involved standard techniques like physical vapor deposition (PVD), photolithography, plasma-enhanced chemical vapor deposition (PECVD), reactive ion etching (RIE), and atomic layer deposition (ALD) for the resistive switching layer.

A novel unsupervised online pattern recognition system was then designed and simulated, incorporating these RRAM synapses and CMOS neurons (specifically, leaky integrate-and-fire circuits). The system consists of 28x28 pre-neurons fully connected to 15 post-neurons via the RRAM crossbar. Post-neurons also have inhibitory connections to each other. The STDP protocol is used to update the RRAM synaptic weights during a training session. When a pre-neuron fires, it sends a signal to the post-neuron. If the post-neuron integrates enough input to reach a threshold, it fires, generating a feedback signal that updates the weights of the synapses connected to it, based on the binary STDP rule.

Datasets

The paper uses the MNIST database of handwritten digits for training and evaluating the pattern recognition system. The 2D digit patterns are converted into a one-dimensional binary input for the pre-neurons.

Key Insights and Ideas

  • Efficient Binary STDP: The paper demonstrates that a binary STDP protocol can be effectively implemented in RRAM devices, leading to efficient learning without requiring complex multilevel resistance states for basic synaptic function.
  • Reduced Training Epochs: By optimizing the STDP protocol and its integration with the RRAM device characteristics, the system shows the potential to learn patterns in fewer epochs compared to previous RRAM-based STDP systems.
  • Unsupervised Online Learning: The proposed system operates in an unsupervised online learning mode, meaning it can learn and adapt without labeled data and can continuously learn new patterns as they are presented.
  • Feasibility of RRAM for Neuromorphic AI: The successful simulation of handwritten digit recognition validates the potential of RRAM-based binary STDP for building practical, energy-efficient neuromorphic computing systems.

Key Results

  • The fabricated RRAM devices exhibit bipolar resistive switching with both abrupt SET (high-resistance state to low-resistance state) and gradual RESET (low-resistance state to high-resistance state) behaviors, enabling multilevel resistance states which contribute to robustness.
  • The binary STDP protocol was experimentally demonstrated, showing that specific temporal overlaps between pre- and post-pulses lead to either long-term potentiation (LTP, switching to LRS) or long-term depression (LTD, switching to HRS).
  • Simulations of the unsupervised online pattern recognition system showed that it could efficiently learn and classify ten handwritten digits from the MNIST database. The weight maps after training indicated that different post-neurons learned to recognize different digits.

Limitations and Future Work

  • Simulation-Based Performance: While the RRAM devices were fabricated and characterized, the pattern recognition system's performance was demonstrated through simulations. Further work would involve implementing and testing the full system in hardware.
  • Idealized Neuron Model: The leaky integrate-and-fire (LIF) neuron model used in the simulation is an idealized representation. Real-world CMOS neuron circuits might introduce additional complexities.
  • Parameter Optimization: The paper mentions that the time span between input and feedback waveforms (2t) needs to be optimized for effective weight updates. Further research could explore more sophisticated methods for parameter tuning and adaptation.
  • Scalability: While the RRAM crossbar array is a step towards scalability, demonstrating the system's performance on larger and more complex datasets would be a valuable next step.
  • Noise and Variability: The paper notes that the RRAM devices show robustness to disturb pulses, but further investigation into the impact of device-to-device variability and noise on the learning performance of the neuromorphic system would be beneficial.

Summary updated

Abstract

A binary spike-time-dependent plasticity (STDP) protocol based on one resistive-switching random access memory (RRAM) device was proposed and experimentally demonstrated in the fabricated RRAM array. Based on the STDP protocol, a novel unsupervised online pattern recognition system including RRAM synapses and CMOS neurons is developed. Our simulations show that the system can efficiently compete the handwritten digits recognition task, which indicates the feasibility of using the RRAM-based binary STDP protocol in neuromorphic computing systems to obtain good performance.
BibTeX
@inproceedings{zhou2017characteristics,
  author = {Zheng Zhou and Chen Liu and Wensheng Shen and Zhen Dong and Zhe Chen and Peng Huang and Lifeng Liu and Xiaoyan Liu and Jinfeng Kang},
  title = {{The Characteristics of Binary Spike-Time-Dependent Plasticity in HfO2-Based RRAM and Applications for Pattern Recognition}},
  booktitle = {Nanoscale},
  year = {2017},
  doi = {10.1186/s11671-017-2023-y},
}
Figure