.png){kind=icon}
.png){kind=icon}
Return
Substantially enhanced reservoir computing with multiple-input mode
.png){kind=icon}
PRE
AI
DOI：10.1088/2631-8695/ae6f79.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Reservoir computing (RC) is a special class of recurrent neural network models. The primary difference of RC from deep learning is that only the connections between the reservoir and the output layer are trainable, and the training requires much less computation than the deep learning. Given that the reservoir topology is fixed as a premise, the input layer is only the tunable section for performance enhancement. To this end, so-called multiple-input mode is proposed for the first time which is an array of identical input data is individually launched in parallel into virtually stacked reservoirs. This contrasts to conventional single-input mode where the single input data is launched to a reservoir. With the proposed multiple-input mode we will thoroughly examine a variety of reservoir topologies, including simplified reservoir topologies such Delay Line Reservoir (DLR), DLR with feedback connections (DLRB), and Simple Cycle Reservoir (SCR), in addition to the conventional RC topology. Benchmark tests of both prediction and classification tasks using time-series datasets are conducted. In the prediction task of Mackey-Glass chaos, it is demonstrated that with the multiple-input simplified topologies such as DLR and SCR can reduce the mean squared error by approximately one order of magnitude, compared with conventional RC. On the other hand, in the classification of six human motions, a third option of input mode, the partial multiple-input mode is introduced. In this mode, some classes use single-inputs, while the rest of classes use multiple-input mode. The benchmark test demonstrates that the classification accuracy of simplified topologies with the partial multiple-input mode substantially improves, compared with that of conventional RC.
Keywords:
reservoir computing
prediction and classification
multiple-input mode
simplified reservoir topologies
virtually stacked reservoir
Journal
E
IF:
1.6
Papers:
2.1K
Citations:
0