loading page

ViterbiNet Transformed
  • Gil Zukerman ,
  • Idan Czuckermann ,
  • Raja Giryes
Gil Zukerman
Tel aviv University

Corresponding Author:[email protected]

Author Profile
Idan Czuckermann
Author Profile
Raja Giryes
Author Profile


The article focuses on the problem of detecting transmitted data over an additive Gaussian noise channel. We propose an integration of various deep-learning architectures with the classical Viterbi algorithm for channel estimation and equalization, to improve the estimation performance.
The goal is to design communication schemes that are oblivious to the channel coefficients.
The study utilizes various deep-learning architectures to address the challenges of signal detection and data reconstruction over AWGN channels in communication systems, by learning the log-likelihood ratio (LLR) /priors of the channel as a function of the channel memory. We draw inspiration from the ViterbiNetmodel-based architecture and explore the implementation of novel deep neural network (DNN) architectures for channel estimation. These architectures include Convolutional Neural Networks (CNN), Transformer Encoder, Residual Convolutional Networks as well as SIONNA (by Nvidia) architecture.
The performance of these architectures was evaluated by bit-error-rate (BER) across different channel conditions by signal-noise-ratio (SNR). The results show an improvement of up to 1dB in signal-to-noise ratio performance in the moderate and high signal-to-noise ratio regions compared to previously proposed learning-based solutions. The experiments demonstrate the potential of DNN model-based architectures for symbol detection and channel equalization in communication systems.
Our work details the methodology used in the experiments, including the details of the DNN architecture implementation, data-set generation, channel model, training process, evaluation metrics, and semi-supervised online training method. The experiments demonstrate the performance of the proposed architectures and compare them to the classical Viterbi algorithm (with known CSI RS) and online Viterbi-Net pervious deep-learning-based architecture.
In conclusion, the article presents the integration of different model-based deep-learning architectures into the Viterbi algorithm for signal reconstruction of transmitted data over an additive Gaussian noise channel. The proposed architectures show promising results in improving the signal-to-noise ratio performance compared to existing learning-based solutions. These findings contribute to the advancement of communication systems and highlight the potential for further improvements in performance and robustness.