In-situ Learning in Hardware Compatible Multi-layer Memristive Spiking Neural Network

Jiwei Li,Hui Xu,Sheng-Yang Sun,Nan Li,Qingjiang Li,Zhiwei Li,Haijun Liu
DOI: https://doi.org/10.1109/tcds.2021.3049487
IF: 4.546
2021-01-01
IEEE Transactions on Cognitive and Developmental Systems
Abstract:As one of the most promising methods in the next generation of neuromorphic systems, memristor-based spiking neural networks (SNNs) show great advantages in terms of power efficiency, integration density, and biological plausibility. However, because of the nondifferentiability of discrete spikes, it is difficult to train SNNs with gradient descent and error backpropagation online. In this article, we propose an improved training algorithm for multilayer memristive SNN (MSNN) with three methods spontaneously, supporting in situ learning in hardware: 1) temporal order encoding is applied to generate different pulse trains in neurons; 2) a simplified homeostasis is realized by the activation state and refractory period to regulate hidden neurons spontaneously; and 3) spiking-timing-dependent plasticity (STDP) in memristive synapses is adopted to update weights in situ. Correspondingly, we provide a circuitry example and verify it in LTSPICE. Then the MSNN is benchmarked with the MNIST data set and analyzed with visualization methods, showing better recognition accuracy (95.15%) than existing SNNs with comparable scales and bio-inspired learning rules. We also consider some nonideal effects in memristor crossbar array and peripheral circuits. Evaluation results show that the proposed MSNN is robust to finite resolution, circuit noise and writing noise; and larger network scale will help the MSNN alleviate the negative impacts of other nonideal factors, including yield and device-to-device variation. Moreover, the energy efficiency of a MSNN system is estimated to achieve 7.6TOPS/W, showing great potential in low-power applications.
robotics,computer science, artificial intelligence,neurosciences
What problem does this paper attempt to address?
The paper mainly aims to address the following issues: ### Core Issues - **Challenges in online training of multi-layer memristor-based Spiking Neural Networks (SNNs)**: Due to the non-differentiability of spike signals, traditional gradient descent and error backpropagation methods are difficult to directly apply to online training of memristor-based SNNs. ### Specific Objectives - **Propose an improved training algorithm**: To solve the problem of online training of multi-layer memristor-based SNNs, the authors propose an improved training algorithm that includes three key methods to support in-situ learning in hardware. - **Temporal coding**: Used to generate different pulse sequences to carry spatiotemporal information. - **Simplified steady-state mechanism**: Spontaneously regulates hidden neurons through activation states and refractory periods. - **Spike-Timing-Dependent Plasticity (STDP)**: Adopts the STDP rule in memristor synapses to update weights in situ. ### Implementation and Verification - **Circuit design and simulation**: A circuit example is provided and verified in LTSpice. - **Performance evaluation**: The proposed MSNN is benchmarked using the MNIST dataset, showing a recognition accuracy of 95.15%, which is superior to existing SNNs of comparable scale with bio-inspired learning rules. - **Robustness analysis**: Non-ideal effects in the memristor crossbar array and peripheral circuits, such as limited resolution, circuit noise, and write noise, are considered, demonstrating that the proposed MSNN has good robustness. - **Energy efficiency estimation**: The energy efficiency of the proposed MSNN system is estimated to reach 7.6 TOPS/W, showing great potential for low-power applications. In summary, this paper aims to overcome the challenges of online training of memristor-based multi-layer SNNs by proposing an improved training algorithm, while also demonstrating its advantages in terms of performance and robustness.