Quantification stochastic configuration networks with incremental encoding

Abstract

In resource-constrained industrial scene, the application of neural networks is a challenge due to the requirement for powerful high-performance computing devices to handle large amounts of floating-point data. The paper proposes a quantified stochastic configuration network model called Stochastic Configuration Networks with Incremental Encoding (SCN-IE), aiming to improve the operating efficiency of the model. To quantize the model, a novel feature encoding is developed to convert the input data into bit vectors. The characteristic of this model is that its hidden layer inputs and weights are represented in the form of bit vectors. We use basic bit logic operations to effectively calculate the output of the hidden layer, achieving lightweight computation. In addition, the stochastic configuration algorithm is used to solve the approximation problem of the model. The results demonstrate that SCN-IE exhibits powerful real-time reasoning capabilities compared to SCN and IRVFLN, and it holds great potential for application on resource-constrained devices.