Scholarlink

<p>Reservoir computing (RC) has been widely applied in the study of nonlinear dynamical systems, serving as a model-free prediction method. Its performance strongly depends on the choice of hyperparameters, among which the network topology is the most complicated factor. In this study, we investigate the influence of three random networks on RC’s memory and prediction capacity, including the Watts–Strogatz (WS) small world network, the Barabási–Albert (BA) scale-free network, and the Erdős–Rényi (ER) random network. It is shown that BA networks exhibit superior memory capacity across a wide range of spectral radius in RC. This advantage is attributed to the hub nodes, which serve as central units for efficient information integration. By embedding hub nodes into ER networks or weakening them in BA networks, we observe substantial changes in both memory capacity and nonlinear prediction accuracy. This finding aligns with biological brain network models, where a hub-centric organization, similar to BA networks, demonstrates better memory capacity compared to ER networks. Our results position the hub nodes as key orchestrators of memory and prediction in both reservoir computing and brain network models, offering valuable insights for the design of next-generation adaptive, high-performance neuromorphic 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

Hub nodes integrate memory and prediction in reservoir computing: insights from graphs to brains

  • 期刊论文

  • 发表日期：

    2026年2月25日

查看全文
  • Minglong He,
  • Qing Li,
  • Zixiang Yan,
  • Jian Gao,
  • Jinghua Xiao

摘要

Reservoir computing (RC) has been widely applied in the study of nonlinear dynamical systems, serving as a model-free prediction method. Its performance strongly depends on the choice of hyperparameters, among which the network topology is the most complicated factor. In this study, we investigate the influence of three random networks on RC’s memory and prediction capacity, including the Watts–Strogatz (WS) small world network, the Barabási–Albert (BA) scale-free network, and the Erdős–Rényi (ER) random network. It is shown that BA networks exhibit superior memory capacity across a wide range of spectral radius in RC. This advantage is attributed to the hub nodes, which serve as central units for efficient information integration. By embedding hub nodes into ER networks or weakening them in BA networks, we observe substantial changes in both memory capacity and nonlinear prediction accuracy. This finding aligns with biological brain network models, where a hub-centric organization, similar to BA networks, demonstrates better memory capacity compared to ER networks. Our results position the hub nodes as key orchestrators of memory and prediction in both reservoir computing and brain network models, offering valuable insights for the design of next-generation adaptive, high-performance neuromorphic computing systems.