Meta-ViterbiNet: Online Meta-Learned Viterbi Equalization for Non-Stationary Channels

Abstract

Deep neural networks (DNNs) based digital receivers can potentially operate in complex environments. However, the dynamic nature of communication channels implies that in some scenarios, DNN-based receivers should be periodically retrained in order to track temporal variations in the channel conditions. To this aim, frequent transmissions of lengthy pilot sequences are generally required, at the cost of substantial overhead. In this work we propose a DNN-aided symbol detector, Meta-ViterbiNet, that tracks channel variations with reduced overhead by integrating three complementary techniques: 1) We leverage domain knowledge to implement a model-based/data-driven equalizer, ViterbiNet, that operates with a relatively small number of trainable parameters; 2) We tailor a meta-learning procedure to the symbol detection problem, optimizing the hyperparameters of the learning algorithm to facilitate rapid online adaptation; and 3) We adopt a decision-directed approach based on coded communications to enable online training with short-length pilot blocks. Numerical results demonstrate that Meta-ViterbiNet operates accurately in rapidly-varying channels, outperforming the previous best approach, based on ViterbiNet or conventional recurrent neural networks without meta-learning, by a margin of up to 0.6dB in bit error rate in various challenging scenarios. Index terms-Viterbi algorithm, meta-learning. I. INTRODUCTIONDeep learning systems have demonstrated unprecedented success in various applications, ranging from computer vision to natural language processing, and recently also digital communications and receiver design [1]- [4]. While traditional receiver algorithms are model-based, relying on mathematical modeling of the signal transmission, propagation, and reception, deep neural networks (DNNs) are model-agnostic, and are trained from data. DNN-aided receivers can operate efficiently in scenarios where the channel model is unknown, highly complex, or difficult to optimize for [5].Despite its potential in implementing digital receivers [6], [7], deep learning solutions are subject to several challenges that limit their applicability in important communication

Deep-Learning-Aided Distributed Clock Synchronization for Wireless Networks

“…Nonetheless, in a dynamic network scenarios with highly mobile nodes, it is likely that the nodes may need to retrain their local models whenever the topology changes considerably from the one used during its training. We expect training schemes designed for facilitating online re-training in rapidly time-varying environments by, e.g., leveraging data from past topologies to predict future variations as in [41], [42]; however, we leave these extensions of DASA for future work.…”

Online Meta-Learning For Hybrid Model-Based Deep Receivers

“…a movement of one or several of the communicating entities. We do not impose a specific model on the channel observed in each block, representing it by the generic conditional distribution in (1), which can take a complex and possibly intractable form. Two important special cases of this channel model include:…”

“…Deep learning systems have demonstrated unprecedented success in various applications, ranging from computer vision to natural language processing, and recently also in digital communication and receiver design [2]- [5]. While traditional receiver algorithms are channel-model-based, relying on mathematical modeling [6] of the signal transmission, propagation, and reception, Parts of this work were presented at the 2021 IEEE International Conference on Communications as the paper [1]. This project has received funding from the Israeli 5G-WIN consertium, the European Union's Horizon 2020 research and innovation program under grants No.…”

Predicting Flat-Fading Channels via Meta-Learned Closed-Form Linear Filters and Equilibrium Propagation

“…, ỹte F ] stacks vertically the L te × 1 vectors {ỹ te f } F f =1 . Unlike standard meta-learning algorithms used by most papers on communications [9], [10], [11], [12], [13], [16], such as model-agnostic meta-learning (MAML), the proposed meta-learning procedure is not based on iterative processing, and it does not require the tuning of hyperparameters such as a learning rate schedule. We will discuss below how to set the only free hyperparameter λ.…”

“…Then, an online algorithm is proposed based on gradient descent and equilibrium propagation (EP) [7], [8]. Previous applications of meta-learning to communication systems include demodulation [9], [10], channel equalization [11], encoding/decoding [12], [13], [14], MIMO detection [15], beamforming [16], [17], and resource allocation [18].…”