Robust Multi-feature Extreme Learning Machine

Abstract

Extreme learning machine as efficiently and effectively single-hidden layer feedforward neural network is widely applied in pattern classification, feature extraction, data prediction and so on. However, it is difficult that ELM deal with multi-feature data directly, because each feature of multi-feature data has a specific statistical property. Therefore, we proposed a robust multi-feature model based on ELM, namely, multi-feature ELM (MFELM). In particular, MFELM builds models to each single feature, and optimize the combinatorial coefficient of multi-feature by iterating. Obtain the robust output weight and the most reasonable combinatorial coefficient that achieve the minimum of training errors sum in different features. In order to avoid the model degrade into single feature and to enhance the robustness of multi-feature, we induce the higher order combinatorial coefficient. Moreover, MFELM is developed to kernel method, and propose robust kernel multi-feature ELM (MFKELM), which solve the problem that different dimensions lead to operate difficultly. MFELM and MFKELM fully explore the complementary property of multi-feature to improve the recognition accuracy, and retain the robustness to multi-feature data. We demonstrate the performance of proposed in image, text, manual multi-feature datasets.