⚡ electrical engineering

LightAMCNet: An Under-70k-Parameter Lightweight Automatic Modulation Recognition Network for Edge Spectrum Intelligence

This paper introduces LightAMCNet, an ultra-lightweight one-dimensional convolutional neural network with only 68.9k parameters that achieves superior accuracy-efficiency trade-offs for real-time automatic modulation recognition on resource-constrained edge devices by utilizing a purely temporal convolutional backbone while avoiding complex architectural components.

Original authors: Cheng Quan, Zhongsheng Guo, Weiyong Sun

Published 2026-07-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Cheng Quan, Zhongsheng Guo, Weiyong Sun

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to identify a song just by listening to a short, 128-second clip of static and noise. In the world of wireless communication, this is exactly what an Automatic Modulation Classification (AMC) system does. It listens to radio signals and figures out what "language" (modulation type) the signal is speaking, without knowing anything about the sender beforehand. This is crucial for things like cognitive radios and smart spectrum management, where devices need to understand their radio environment instantly.

The problem is that most current "experts" at this task are like over-engineered supercomputers. They are huge, heavy, and require massive amounts of memory and power to run. If you try to put one of these giants on a small, battery-powered drone or a tiny sensor, it simply won't fit or will drain the battery instantly.

The Solution: LightAMCNet

The authors of this paper, Cheng Quan and his team, decided to build a lightweight expert. They created a new neural network called LightAMCNet.

Think of LightAMCNet as a sleek, minimalist detective compared to the bloated, suit-wearing detectives of the past.

The Size: While other models might have hundreds of thousands or even millions of "brain cells" (parameters), LightAMCNet has only 68,900. That's less than 1/6th the size of the previous best model (MCLDNN).
The Design: Instead of using fancy, complex tools like "attention mechanisms" (which let the model focus on specific parts of the signal) or "frequency-domain branches" (which look at the signal's frequency spectrum), LightAMCNet uses a pure, simple time-based approach.
- Analogy: Imagine trying to recognize a person's voice. Some methods try to analyze the sound waves, the pitch, the rhythm, and the background noise all at once using complex machinery. LightAMCNet just listens to the rhythm and the shape of the sound waves over time, step-by-step, using a very simple, straight-line approach. It deliberately avoids the "fancy gadgets" to stay small and fast.

How It Works

The network takes a raw stream of radio data (called I/Q sequences) and runs it through four simple layers of "temporal convolution."

The Stem: It starts by grabbing the raw signal.
The Backbone: It passes the signal through four blocks that get progressively "wider" (looking at more data at once), but they remain very simple. They don't use complex shortcuts or special attention filters.
The Finish: At the end, it uses two different "pooling" methods (like taking an average of the whole signal and also picking out the loudest, most important spikes) to make a final guess.

The Results: Small but Mighty

The team tested their model on two famous radio signal datasets (RadioML 2016.10a and 10b). Here is what happened:

Accuracy: Despite being tiny, LightAMCNet actually beat the much larger MCLDNN model. On the first dataset, it was 3.38% more accurate while using only 16.9% of the memory.
Speed: Because it is so simple, it is incredibly fast. On a high-end graphics card, it can process about 198,000 signals per second. That's roughly three times faster than the larger model it beat.
Efficiency: The researchers played with the "width" of the network (making it narrower or wider). They found that even a very narrow version (with only 20k parameters) could still get the job done with decent accuracy, proving the design is flexible.

Why This Matters (According to the Paper)

The paper argues that you don't need a massive, complex brain to recognize radio signals effectively. A carefully designed, simple, time-based approach is enough to get high accuracy while keeping the model small enough to fit on edge devices (like sensors or small radios).

The authors are careful to note that this was tested on standard computer datasets and a powerful desktop graphics card. They haven't yet tested it on tiny microchips or in real-world, messy radio environments with interference, but the results suggest that a "simple is better" approach could be the key to putting smart spectrum intelligence into small, portable devices in the future.

In short: The paper proves that by stripping away the complex, heavy machinery of modern AI and sticking to a simple, pure time-based design, you can create a radio signal classifier that is tiny, fast, and surprisingly accurate.

Technical Summary: LightAMCNet

Problem Statement

Automatic Modulation Classification (AMC) is a critical enabler for cognitive radio, spectrum monitoring, and edge spectrum intelligence in next-generation mobile networks. While deep learning-based AMC methods have achieved high recognition performance, many existing models rely on complex architectures containing hundreds of thousands to millions of trainable parameters. These models often incorporate attention mechanisms, Long Short-Term Memory (LSTM) layers, multi-branch time-frequency processing, or specialized components like squeeze-and-excitation modules. Such complexity increases memory footprint, implementation complexity, and inference costs, rendering them unsuitable for resource-constrained wireless nodes, unmanned platforms, and software-defined radio devices. Furthermore, existing "lightweight" methods often still exceed 100k parameters or require specialized operations that complicate edge deployment.

Methodology

The paper proposes LightAMCNet, a compact one-dimensional convolutional neural network designed to operate with fewer than 70,000 parameters (specifically 68.9k). The architecture is built on a "pure temporal convolutional backbone" and deliberately avoids several common deep learning design patterns to ensure simplicity and deployability.

Core Architectural Choices

Pure Temporal Convolution: The network processes raw In-phase/Quadrature (I/Q) sequences directly in the time domain. It avoids frequency-domain branches (e.g., Fourier transforms, Short-Time Fourier Transforms) and frequency-domain gating.
Simplified Building Blocks: The core component is the Plain Temporal Block (PTB), a residual block that:
- Uses standard 1D convolutions (kernel sizes 5 and 3) instead of depthwise separable or grouped convolutions.
- Excludes channel attention mechanisms (e.g., SE, ECA, CBAM) and self-attention modules.
- Avoids inverted residual blocks (MBConv) and channel expansion mechanisms.
Network Structure:
- Stem: A single 1D convolution (kernel size 7) mapping 2-channel I/Q input to initial feature channels.
- Backbone: Four stages of PTBs with gradually increasing channel dimensions ( $[16, 24, 32, 48, 64] \times w$ ). Stride-2 downsampling reduces temporal resolution from 128 to 16 samples.
- Head: A dual-pooling layer (parallel adaptive average and max pooling) followed by a two-layer classifier with SiLU activation and dropout.
Scalability: A width multiplier ( $w \in \{0.5, 0.75, 1.0, 1.5\}$ ) allows for trading accuracy against model size. The default configuration uses $w=1.0$ .

Key Contributions

Model Development: Introduction of LightAMCNet, an AMC model with only 68.9k trainable parameters, significantly smaller than baselines like MCLDNN (406.5k) and SCNN (104.1k).
Performance Validation: Demonstration that a carefully configured plain temporal convolutional architecture can achieve competitive accuracy on standard benchmarks (RadioML 2016.10a and 2016.10b) without complex components.
Ablation Analysis: Systematic evaluation of the width multiplier and dropout rate to characterize the accuracy-efficiency trade-off.
Efficiency Metrics: Comprehensive evaluation of computational cost, including parameters, Multiply-Accumulate Operations (MACs), latency, and batch throughput.

Experimental Results

Experiments were conducted on the RadioML 2016.10a and RadioML 2016.10b datasets using an NVIDIA RTX 4070 SUPER GPU.

Accuracy:
- RadioML 2016.10a: LightAMCNet achieved 61.92% accuracy, outperforming MCLDNN (58.54%) by 3.38 percentage points while using only 16.9% of MCLDNN's parameters.
- RadioML 2016.10b: LightAMCNet achieved 64.56% accuracy, slightly surpassing MCLDNN (64.28%).
- Comparison: Both LightAMCNet and MCLDNN significantly outperformed the SCNN baseline (53.35% and 53.84%, respectively).
Efficiency:
- Throughput: LightAMCNet achieved a batch inference throughput of approximately 198,000 inferences per second (batch size 256), roughly three times that of MCLDNN (65.2k).
- Latency: Single-sample latency was 1.23 ms, comparable to MCLDNN (1.33 ms) and slightly higher than SCNN (0.26 ms).
- Memory: The 68.9k parameters require approximately 276 KB of storage (32-bit float), compared to 1.63 MB for MCLDNN.
Ablation Findings:
- Accuracy saturated around a width multiplier of $w=1.0$ ; increasing to $w=1.5$ yielded no improvement and exceeded the 100k parameter budget.
- Even at $w=0.5$ (20.2k parameters), the model maintained 61.24% accuracy on 2016.10a.
- A dropout rate of 0.15 provided a stable balance, though 0.05 yielded slightly higher accuracy on 2016.10b.

Significance and Claims

The paper posits that carefully designed plain temporal convolutions can provide a favorable accuracy-efficiency trade-off for real-time AMC in edge wireless systems. The significance of LightAMCNet lies in its ability to:

Democratize Edge Deployment: By reducing the parameter count to under 70k, the model is suitable for storage and execution on low-power wireless nodes and embedded accelerators where memory and compute are limited.
Simplify Implementation: By avoiding specialized operations (attention, frequency transforms, depthwise separable convolutions), the architecture is easier to implement, reproduce, and optimize (e.g., via quantization) on hardware.
Validate Simplicity: The results challenge the assumption that complex, multi-branch, or attention-heavy architectures are strictly necessary for high-performance AMC, suggesting that sufficient temporal depth and dual-pooling strategies are more critical for I/Q signal representation in this context.

The authors acknowledge limitations, noting that evaluations were performed on a desktop GPU rather than microcontrollers or FPGAs, and the model currently addresses single-label classification without handling overlapping signals. Future work is proposed to explore knowledge distillation, quantization-aware training, and deployment on actual embedded wireless hardware.