基于多反馈环结构提高硬件储备池记忆能力

Improving Memory Capacity of Hardware Reservoir Computing by Multiple Feedback Loops

针对单反馈环硬件储备池记忆能力不足的问题,提出一种基于多反馈环结构提高储备池记忆能力的方法.通过增加反馈环将更早输入信号产生的响应重新注入回储备池,使更早期的输入信号也影响储备池的内部动态,从而增强储备池的记忆能力.仿真研究了单个和多个反馈环结构的储备池的记忆能力及其在需要更长记忆能力的NARMA30任务中的预测性能.结果表明,把反馈环增加到10个,虚节点数为50时储备池的记忆能力由单反馈环的18.2提高到40.2.仅用两个反馈环,虚节点数为1000时NARMA30预测的归一化均方根误差可从单反馈环的0.27降到0.09.说明通过合理设置多个反馈环的参数,可以设计出任务需要的特定记忆能力,部分解决了储备池的适应性问题.

In order to improve the memory capacity of hardware reservoir computing （HRC） based on a single feed-back loop, a HRC scheme based on multiple feedback loops is proposed. Adding extra loops can feedback the responses stimulated by past input signals into the reservoir to increase the HRC memory capacity. Investigations on the performances of HRCs based on a single and multiple feedback loops were carried out by numerical simulation for the memory capacity and a NARMA30 task which needs a long memory capacity. Results show that the HRC memory capacity is increased to 40.2 at ten loops from 18.2 at a single loop when the node number is 50. For the NARMA30 task, the Normalized Root Mean Square Error is decreased to 0. 09 at two loops from 0. 27 at a single loop when the node number is 1000. Therefore, specific memory capacity needed by a task can be designed through setting the parameters of feedback loops, partially solving the a- daptive problem of reservoir computing.