Interview

Abstract

Xiaofeng Yang Xiaofeng Yang, of Yulin Normal University in China, talks to us about her group's submission “Multi-class import vector machine for transmit antenna selection in MIMO systems” page 62. I am currently an Associate Professor with the School of Physics and Telecommunication Engineering, Yulin Normal University, China. My research interests focus is machine learning algorithm design for wireless communications applications, such as wireless positioning and wireless network optimisation. Machine learning algorithms are able to learn a function mapping from input features to desired output, even there is a complicated nonlinear relation between the two. The major advantage of machine learning methods is that they are non-parametric approaches, without the requirement of a statistical characterisation of channels or received waveforms, whilst achieving better accuracy than conventional methods. The main contributions of our resent research were developing learning based non-line-of-sight (NLOS) identification and NLOS mitigation algorithms to alleviate the severe degradation of wireless positioning accuracy in dense multipath environment. Multiple-input multiple-output (MIMO) technology is one of the key technologies in 4G/5G wireless communications systems. By employing multiple antennas at the transmitter and/or receiver, MIMO systems achieve better spatial diversity and reap higher multiplexing gain, but the hardware complexity and cost is considerably high. One promising technique that addresses the issue is antenna selection, the idea behind which is to adopt a reasonable subset of available antennas at the transmitter and/or receiver, and to reduce the number of RF chains to the same size as selected antennas, therefore making the deployment of MIMO systems much more feasible. Conventional antenna selection methods require an optimisation-driven decision, which leaves much space for improvement in either complexity or precision. Our Letter puts forth a novel, learning-based antenna selection approach. A multi-class classifier, import vector machine (IVM), maximises the average received signal-to-noise ratio (SNR). IVM is an excellent classifier with considerable merits: it has outstanding classification performance, it is computationally efficient with very few import vectors and it allows for probabilistic interpretation of the classification results. There are very few publications in the literature that apply IVM to the region of wireless communications. To the best of our knowledge, our Letter is the first time to explore the advantages of IVM as a multi-class classifier in the context of antenna selection for received SNR maximisation. Our simulation results prove that IVM outperforms the conventional optimisation-driven algorithm and the state-of-the-art learning based scheme of support vector machine (SVM) in terms of average received SNR performance with feasible complexity and sparsity. Therefore, IVM is a very efficient antenna selection approach for MIMO systems and has very promising application prospect in practice. The antenna selection problem can be translated into a multi-class classification learning task: to project an input sample to one of the available antenna subsets according to certain criteria. IVM was initially introduced as a binary classifier in the literature. The main challenge of our research was to generalise IVM to the case of multi-class classification. Our Letter was the first attempt to provide explicit formulation and explanation. The basic idea of our multi-class classifier IVM is that it estimates the probability that the input sample belongs to each class, and categorises the sample to the class with the biggest likelihood. The other challenge was to design an appropriate input feature for the multi-class classifier IVM. We developed a novel input feature, the Euclidean norm of channel matrix, which characterises the salient property to identify the optimal antenna subsets. As for the theoretical research, deep learning is a new breed of machine learning technique, which is gaining much popularity in academia and industry. One of the promises of deep learning is replacing handcrafted features with efficient algorithms for unsupervised or semi-supervised feature learning and hierarchical feature extraction. Meanwhile, the research field of machine learning has expanded from pure theory to chip implementation in recent years. There are some machine learning processors budding in industry, providing a massive uplift in efficiency compared to CPUs, GPUs and DSPs through efficient convolution, sparsity and compression. In the near future, we expect to see more innovative artificial intelligent applications based on machine learning processors for a wide array of market segments including mobile, automotive and infrastructure, to name a few.