Multi-subspace RBFNN driven by features correlation learning

For the radial basis function neural network (RBFNN), the centers of the kernels and the network weight are critical to the network performance. The expectation maximization (EM) algorithm can learn the network parameters in RBFNN adaptively. However, high-dimensional samples often have complex distributions in the feature space. In this case, the kernels learned by the EM algorithm may be inaccurate or the algorithm cannot converge. To address these problems, this paper proposes a multi-subspace RBFNN (MS-RBFNN) based on features correlation learning. Inspired by the two-dimensional convolutional neural network sliding window operation to extract local information, this paper uses the correlated features to construct feature subsets with local characteristics. In each subspace, a multi-layer RBFNN is used to extract the local distribution response characteristics of samples. Finally, the distributed response features in multi subspaces are combined to perform classification tasks. This method can use more potential local characteristics of vector features to learn more information about the data. At the same time, The constructed low-dimensional subspaces make the network easier to train and use. Experiment results show that compared with some existing algorithms, the model proposed in this paper has a higher classification accuracy than the state-of-the-art method.