Scholarlink

<p>Memristor-based neuromorphic computing systems promise brain-like efficiency but face challenges in training deep networks due to the number of parameters. Reservoir computing (RC) offers a compelling alternative by employing a reservoir layer to preprocess temporal data, thereby reducing network complexity and training costs. This work explores the implementation of an RC system utilizing organic polyaniline (PANI) memristors, which are well-suited for this role due to their inherent short-term memory and tunable switching dynamics. We investigate the critical, yet understudied, relationship between the reservoir’s information capacity (number of bits of information for processing) and system performance. For this purpose, the RC system is benchmarked on the MNIST handwritten digit recognition task. Our results demonstrate that high information capacity allows a significant reduction in the number of synaptic weights in the subsequent readout layer with only a marginal decrease in classification accuracy. Additionally the impact of the way of image pre-processing is analysed. We offer an optimal capacity value and image pre-processing that minimizes system parameters while maintaining high performance, providing key design insight for efficient memristor-based reservoir computing systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Polyaniline Memristor-based Physical Reservoir Computing with Different Reservoir’s Information Capacity

  • 期刊论文

  • 发表日期：

    2026年4月15日

查看全文
  • Maria V. Serenko,
  • Aleksey V. Serenko,
  • Andrey V. Emelyanov,
  • Gang Liu,
  • Vyacheslav A. Demin

摘要

Memristor-based neuromorphic computing systems promise brain-like efficiency but face challenges in training deep networks due to the number of parameters. Reservoir computing (RC) offers a compelling alternative by employing a reservoir layer to preprocess temporal data, thereby reducing network complexity and training costs. This work explores the implementation of an RC system utilizing organic polyaniline (PANI) memristors, which are well-suited for this role due to their inherent short-term memory and tunable switching dynamics. We investigate the critical, yet understudied, relationship between the reservoir’s information capacity (number of bits of information for processing) and system performance. For this purpose, the RC system is benchmarked on the MNIST handwritten digit recognition task. Our results demonstrate that high information capacity allows a significant reduction in the number of synaptic weights in the subsequent readout layer with only a marginal decrease in classification accuracy. Additionally the impact of the way of image pre-processing is analysed. We offer an optimal capacity value and image pre-processing that minimizes system parameters while maintaining high performance, providing key design insight for efficient memristor-based reservoir computing systems.