Robust MIMO Detection using Hypernetworks with Learned Regularizers

Abstract

Optimal symbol detection in multiple-input multiple-output (MIMO) systems is known to be an NP-hard problem. Recently, there has been a growing interest to get reasonably close to the optimal solution using neural networks while keeping the computational complexity in check. However, existing work based on deep learning shows that it is difficult to design a generic network that works well for a variety of channels. In this work, we propose a method that tries to strike a balance between symbol error rate (SER) performance and generality of channels. Our method is based on hypernetworks that generate the parameters of a neural network-based detector that works well on a specific channel. We propose a general framework by regularizing the training of the hypernetwork with some pre-trained instances of the channel-specific method. Through numerical experiments, we show that our proposed method yields high performance for a set of prespecified channel realizations while generalizing well to all channels drawn from a specific distribution.

Annealed Langevin Dynamics for Massive MIMO Detection

“…Hence, it is unsuitable for real-time applications. However, this issue can be alleviated by using HyperNetworks [12], [13].…”

“…In the new forward model in (12), the likelihood is given by p(y|x l , H) = p(z − Hn l |x l ), which is not Gaussian: although p(n l ) is a Gaussian distribution, after conditioning on xl the conditional distribution p(n l |x l ) is no longer Gaussian. To understand this, notice that (11), p(n l |x l ) ∝ p(x l |n l )p(n l ) is a mixture of Gaussians weighted by the distribution of x.…”

“…Before moving to this explanation, and in order to get a tractable expression, we have to rely on the singular value decomposition (SVD) of the channel matrix given by H = UΣV as well as in the spectral representation of xl , y, z and n l defined as χl = V xl , η = U y, ζ = U z and ν l = V n l . Given the spectral representation, the model in (12) can be rewritten as…”

“…They have shown promising results, with a good balance between performance and complexity. They can be roughly classified into data-driven methods, like DetNet [9], and model-driven methods, like RE-MIMO [10], OAMPNet [11], and hyperMIMO [12], [13]. The main issue with all of the learning-based detectors is their dependency on a particular configuration of the communication system as they are trained for a specific modulation scheme and, with the exception of RE-MIMO, for a given system size.…”