GenELM: Generative Extreme Learning Machine feature representation

Extreme Learning Machine (ELM) feature representation has been drawing increasing attention, and most of the previous works devoted to learning discriminative features. However, we argue that such kind of features suffer from “categories bias” in target detection tasks, where the scope of the negatives (i.e., backgrounds) is naturally broader than that of the positives (i.e., targets). Technically, the learning of discriminative features aims to maximize positive-negative difference. In this case, the interclass-difference criteria for feature mapping and feature matching would increase model complexity while degrading detection robustness. To address these issues, we construct a generative feature learning framework in this paper. Unlike the discriminative feature representation, generative features tend to cluster the positives from negatives, and in this way, the unique and separable attributes of the positives can be learnt to facilitate the accurate detection of positives (targets). Practically, by exploiting ELM computational advantages, positives clustering oriented feature learning is implemented with a kernel-label alignment function, which guides random features of positives into clusters, while keeping away from the negatives. On the other hand, the density of positives cluster is estimated using a probabilistic model, and thus the likelihood that test samples belonged to the positive category can be approximately inferred. We experimentally evaluate our framework for target detection on synthetic data and Hyperspectral Imagery (HSI), in which reasonable improvements are observed over conventional ELM-based discriminative methods and backpropagation-based multilayer perceptron (MLP) in terms of detection accuracy or learning efficiency.