Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Learning Models for Physical Layer Communications

Nunzio A. Letizia

TL;DR

This thesis investigates deep learning approaches to physical layer communications, reframing channel capacity, coding, and decoding as data-driven optimization problems. It advances three interwoven streams: (1) copula-based density estimation and segmented generative networks for flexible statistical modeling of channels and dependencies; (2) end-to-end neural schemes for medium modeling, neural decoding (MIND), and capacity-oriented autoencoders that jointly maximize mutual information and decoding reliability; and (3) variational MI estimation (DIME) with derangement strategies to enable stable, scalable MI and capacity estimation across diverse channels, including PLC scenarios. Key innovations include CODINE for copula density estimation, SGN-C/SGN-D for dependence-aware data synthesis, GAN-based channel synthesis with time–frequency noise modeling, the MIND neural decoder, rate- and capacity-driven autoencoders with MI regularization, and a broad family of f-divergence MI estimators (γ-DIME, KL-/GAN-/HD-DIME) with derangement-based sampling. Collectively, these contributions yield practical, data-driven tools for channel synthesis, neural decoding in unknown channels, and capacity-approaching end-to-end communication systems, with demonstrable gains in robustness and performance across AWGN, non-Gaussian, and Rayleigh-like environments, including PLC contexts. The work outlines a cohesive framework for integrating ML with communications theory to approach, estimate, and exploit channel capacity in challenging, real-world media.

Abstract

The increased availability of data and computing resources has enabled researchers to successfully adopt machine learning (ML) techniques and make significant contributions in several engineering areas. ML and in particular deep learning (DL) algorithms have shown to perform better in tasks where a physical bottom-up description of the phenomenon is lacking and/or is mathematically intractable. Indeed, they take advantage of the observations of natural phenomena to automatically acquire knowledge and learn internal relations. Despite the historical model-based mindset, communications engineering recently started shifting the focus towards top-down data-driven learning models, especially in domains such as channel modeling and physical layer design, where in most of the cases no general optimal strategies are known. In this thesis, we aim at solving some fundamental open challenges in physical layer communications exploiting new DL paradigms. In particular, we mathematically formulate, under ML terms, classic problems such as channel capacity and optimal coding-decoding schemes, for any arbitrary communication medium. We design and develop the architecture, algorithm and code necessary to train the equivalent DL model, and finally, we propose novel solutions to long-standing problems in the field.

Deep Learning Models for Physical Layer Communications

TL;DR

This thesis investigates deep learning approaches to physical layer communications, reframing channel capacity, coding, and decoding as data-driven optimization problems. It advances three interwoven streams: (1) copula-based density estimation and segmented generative networks for flexible statistical modeling of channels and dependencies; (2) end-to-end neural schemes for medium modeling, neural decoding (MIND), and capacity-oriented autoencoders that jointly maximize mutual information and decoding reliability; and (3) variational MI estimation (DIME) with derangement strategies to enable stable, scalable MI and capacity estimation across diverse channels, including PLC scenarios. Key innovations include CODINE for copula density estimation, SGN-C/SGN-D for dependence-aware data synthesis, GAN-based channel synthesis with time–frequency noise modeling, the MIND neural decoder, rate- and capacity-driven autoencoders with MI regularization, and a broad family of f-divergence MI estimators (γ-DIME, KL-/GAN-/HD-DIME) with derangement-based sampling. Collectively, these contributions yield practical, data-driven tools for channel synthesis, neural decoding in unknown channels, and capacity-approaching end-to-end communication systems, with demonstrable gains in robustness and performance across AWGN, non-Gaussian, and Rayleigh-like environments, including PLC contexts. The work outlines a cohesive framework for integrating ML with communications theory to approach, estimate, and exploit channel capacity in challenging, real-world media.

Abstract

The increased availability of data and computing resources has enabled researchers to successfully adopt machine learning (ML) techniques and make significant contributions in several engineering areas. ML and in particular deep learning (DL) algorithms have shown to perform better in tasks where a physical bottom-up description of the phenomenon is lacking and/or is mathematically intractable. Indeed, they take advantage of the observations of natural phenomena to automatically acquire knowledge and learn internal relations. Despite the historical model-based mindset, communications engineering recently started shifting the focus towards top-down data-driven learning models, especially in domains such as channel modeling and physical layer design, where in most of the cases no general optimal strategies are known. In this thesis, we aim at solving some fundamental open challenges in physical layer communications exploiting new DL paradigms. In particular, we mathematically formulate, under ML terms, classic problems such as channel capacity and optimal coding-decoding schemes, for any arbitrary communication medium. We design and develop the architecture, algorithm and code necessary to train the equivalent DL model, and finally, we propose novel solutions to long-standing problems in the field.

Paper Structure

This paper contains 189 sections, 29 theorems, 346 equations, 96 figures, 17 tables, 7 algorithms.

Table of Contents

  1. Introduction
  2. Motivation
  3. Scientific contributions
  4. Thesis outline
  5. Publications
  6. Part 1
  7. Deep Learning Fundamentals
  8. Learning theory concepts
  9. Supervised learning
  10. Unsupervised learning
  11. Reinforcement learning
  12. Maximum likelihood estimation
  13. Statistical distances
  14. Machine learning tools
  15. Neural networks
  16. ...and 174 more sections

Key Result

Theorem 3.2.1.1

Let $\mathbf{u} \sim c_U(\mathbf{u})$ be $d$-dimensional samples drawn from the copula density $c_U$. Let $f^*$ be the Fenchel conjugate of $f:\mathbb{R}_+ \to \mathbb{R}$, a convex lower semicontinuous function that satisfies $f(1)=0$ and has derivative $f^{\prime}$. If $\pi_U(\mathbf{u})$ is a mul then where

Figures (96)

  • Figure 1: Different engineering design approaches: a) bottom-up and b) top-down.
  • Figure 2: Flowchart illustrating when to use ML for communication engineering design.
  • Figure 3: Taxonomy of learning models. Deterministic models extract either fixed relationships between input and output (supervised) or patterns of the input (unsupervised). Probabilistic discriminative models use the input to predict either the output (supervised) or the hidden variable causing the input (unsupervised). Probabilistic generative models learn the statistical relationship between either input and output (unsupervised) or input and the hidden variable (supervised). Autoencoders model how to encode the input into the hidden variable, as well as how to decode from the hidden variable to the input.
  • Figure 4: Architecture of a fully connected neural network with two hidden layers.
  • Figure 5: Structure of a convolutional neural network with one convolutional layer.
  • ...and 91 more figures

Theorems & Definitions (51)

  • Theorem 3.2.1.1
  • proof
  • Corollary 3.2.1.1
  • proof
  • Lemma 5.2.2.1
  • proof
  • Lemma 6.3.2.1
  • proof
  • Theorem 7.2.2.1
  • Lemma 7.2.2.1
  • ...and 41 more