For decades, wireless communication and sensing systems have relied on a divide-and-conquer strategy. A receiver does not simply translate raw electromagnetic waves into digital bits in one giant leap. Instead, it passes the incoming signal through a carefully orchestrated pipeline of isolated functional blocks: channel estimation, equalization, interference cancellation, demodulation, and finally, error-correction decoding. Each block is designed using elegant, closed-form mathematical models optimized for that specific sub-task.
However, as we push into the regimes of next-generation communications—characterized by massive MIMO arrays, millimeter-wave (mmWave) frequencies, and Reconfigurable Intelligent Surfaces (RIS)—this modular paradigm is hitting a computational and physical wall. The mathematical models we rely on are becoming too complex to compute within microsecond latency constraints, or they fail to capture the messy, non-linear realities of actual hardware.
The emerging alternative is a paradigm shift toward monolithic, data-driven receiver architectures. Instead of breaking the receiver into discrete mathematical steps, this approach uses a single, end-to-end deep neural network to map raw received signals directly to the final desired output, completely bypassing explicit intermediate steps like channel estimation.
The Pitfalls of Modularity
To understand why monolithic architectures work, we must first look at why the traditional modular approach fails in complex environments. Modularity introduces two hidden costs: error propagation and metric mismatch.
First, consider error propagation. In a traditional receiver, the equalization block assumes that the channel estimation block provided a perfect mathematical map of the environment. If the estimation is even slightly inaccurate, that error compounds as the signal moves down the processing chain. As Linglong Dai et al. point out in Deep Learning for Wireless Communications: An Emerging Interdisciplinary Paradigm, an integrated receiver architecture processes these tasks jointly within a single unified network, allowing the system to better account for non-linear distortions and eliminate the accumulation of errors between separate functional blocks.
Second, modular designs suffer from a fundamental metric mismatch. The channel estimation block is typically optimized to minimize Mean Squared Error (MSE). However, the ultimate goal of a communication system is not to create an accurate mathematical description of the channel; it is to maximize data throughput or minimize the bit error rate. As Kareem M. Attiah et al. observe in Deep Learning for Channel Sensing and Hybrid Precoding in TDD Massive MIMO OFDM Systems, traditional two-step processes are inherently limited by this mismatch because optimizing an intermediate step for MSE does not guarantee an optimal final system throughput.
Bypassing the Channel State Information Bottleneck
The most radical departure of the monolithic architecture is the elimination of explicit Channel State Information (CSI). In traditional systems, knowing the exact mathematical coefficients of the wireless channel matrix is non-negotiable. But in modern architectures like RIS, estimating these high-dimensional matrices requires an unsustainable number of "pilot signals" (calibration overhead), consuming the very bandwidth the system is trying to provide.
Monolithic networks solve this by treating the wireless channel as a latent representation rather than an explicit matrix to be reconstructed.
Traditional Pipeline:
[Raw Pilots] ──> [Channel Estimation (CSI)] ──> [Matrix Inversion] ──> [Beamforming Weights]
Monolithic Pipeline:
[Raw Pilots] ───────────────────> [Neural Network] ───────────────────> [Beamforming Weights]
Instead of calculating intermediate channel matrices, the network learns a direct mapping from raw inputs to final outputs. For example, Tao Jiang et al. demonstrated in Learning to Beamform for Intelligent Reflecting Surface with Implicit Channel Estimate that a neural network can transform raw pilot signals directly into hardware beamforming configurations. By skipping explicit channel reconstruction entirely, the system extracts relevant control features directly from the received signal without wasting overhead on full channel matrix reconstruction.
This direct mapping approach is highly versatile. It can map raw signals directly to transmitted bits, as shown by H. Wang et al. in Deep learning based signal processing and detection for multiple medical devices OFDM systems, allowing receivers to maintain high detection accuracy even when the density of pilot signals is significantly reduced.
Internalizing the Messy Real World
Traditional communication theory relies on clean, mathematically tractable models of noise and distortion—such as Additive White Gaussian Noise (AWGN) or linear fading. Real-world hardware, however, is messy. Power amplifiers introduce non-linear distortions, oscillators drift, and industrial environments feature unpredictable, "spiky" impulsive noise.
Writing an explicit mathematical equation for every hardware imperfection is virtually impossible. Monolithic receivers excel here because they are model-agnostic. By training on raw input-output data pairs, the network implicitly internalizes the physical channel and hardware characteristics.
For instance, Sangwoo Park et al. show in Learning How to Demodulate from Few Pilots via Meta-Learning that framing demodulation as a direct supervised learning task allows a neural network to implicitly account for non-linearities like amplifier distortion directly from data. Similarly, in full-duplex systems where a radio transmits and receives simultaneously on the same frequency, Besma Smida et al. explain in Full-Duplex Wireless for 6G: Progress Brings New Opportunities and Challenges that deep neural networks can handle self-interference cancellation by modeling and mitigating complex non-linear distortions without requiring prior knowledge of the specific distortion model.
When deployed in chaotic environments, such as industrial settings with severe impulsive noise, replacing modular signal processing chains with a single end-to-end network allows for a direct mapping from distorted IQ signals to bitstreams, bypassing separate blocks that frequently fail under non-Gaussian noise distributions, as demonstrated by Bin Wang et al. in An intelligent signal processing method against impulsive noise interference in AIoT.
Why It Works: The Outfielder Analogy
To build intuition for why this works, imagine an outfielder trying to catch a baseball.
A modular, model-based approach to catching the ball would involve measuring the ball's initial velocity, estimating the wind speed, calculating air resistance, and solving differential equations of motion to predict the exact latitude and longitude where the ball will land. The outfielder would then run to that calculated spot. If the wind speed sensor is slightly uncalibrated, the entire calculation fails, and the outfielder misses the ball.
A monolithic, data-driven approach is how humans actually catch baseballs. The outfielder looks at the raw visual feed of the ball and directly adjusts their running speed and direction to keep the angle of the ball constant relative to their eyes. The outfielder does not know the wind speed in miles per hour, nor do they know the ball's trajectory equation. They treat the complex physics of the environment as a latent factor, optimizing purely for the end goal: making contact with the ball.
Monolithic receivers act like that experienced outfielder. By optimizing the entire architecture against a single end-to-end loss function—such as minimizing symbol error rate or maximizing data rate—the neural network learns to ignore irrelevant environmental variations and focuses exclusively on the latent features required to recover the underlying information.
Key Foundational Work
Detection Algorithms for Communication Systems Using Deep Learning (2017) — contributed 1 concept elements to this direction.
Performance of Synthetic Neural Network Classification of Noisy Radar Signals (1988) — contributed 1 concept elements to this direction.