⚡ electrical engineering

MoEMba: A Mamba-based Mixture of Experts for High-Density EMG-based Hand Gesture Recognition

The paper proposes MoEMba, a novel Mamba-based Mixture of Experts framework that integrates channel attention and wavelet feature modulation to effectively capture temporal and spatial dependencies in HD-sEMG signals, thereby significantly improving inter-session and inter-subject hand gesture recognition accuracy.

Original authors: Mehran Shabanpour, Kasra Rad, Sadaf Khademi, Arash Mohammadi

Published 2026-02-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mehran Shabanpour, Kasra Rad, Sadaf Khademi, Arash Mohammadi

Original paper licensed under CC BY 4.0 (http://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 teach a robot to understand your hand gestures so it can control a prosthetic arm or a video game character just by watching your muscles twitch. You use a special "muscle camera" (called HD-sEMG) that has 128 tiny sensors stuck to your skin to listen to the electrical signals your muscles make.

The problem? Your muscles are inconsistent.

If you try to make a "thumbs up" gesture today, the signal looks one way. If you try it tomorrow, or if you're tired, or if you moved the sensors slightly, the signal changes completely. It's like trying to recognize a friend's voice over a phone line that keeps changing its static noise. Most computer programs get confused by this and fail when you try them on a new day or a new person.

This paper introduces a new AI system called MoEMba (a clever play on "Mamba," the snake, and "Mixture of Experts") that solves this problem. Here is how it works, using simple analogies:

1. The "Musical Ear" (Wavelet Transform)

Imagine listening to a complex song. A regular computer might just hear the volume. But MoEMba has "musical ears." It uses something called a Wavelet Transform.

The Analogy: Think of the muscle signal as a messy orchestra. MoEMba doesn't just listen to the whole noise; it separates the music into different instruments (the bass, the drums, the violins) and different speeds (slow rhythms vs. fast beats).
Why it helps: This lets the AI see the "shape" of the gesture in both time and frequency, making it easier to spot the difference between a "fist" and a "point," even if the signal is a bit fuzzy.

2. The "Team of Specialists" (Mixture of Experts)

Instead of one giant brain trying to memorize every possible muscle signal, MoEMba hires a team of specialists.

The Analogy: Imagine a hospital emergency room. You don't have one doctor who knows everything about every disease. Instead, you have a heart doctor, a brain doctor, and a bone doctor. When a patient arrives, a smart Triage Nurse (the Gating Network) looks at the symptoms and instantly decides: "This is a heart issue, send them to Dr. Heart," or "This is a bone issue, send them to Dr. Bone."
How it works: In MoEMba, the "doctors" are called Mamba Experts. One expert might be great at recognizing fast, jerky movements, while another is great at slow, smooth ones. The Triage Nurse routes the data to the right expert. This makes the system much smarter and faster than a single giant brain.

3. The "Snake that Remembers" (Mamba Architecture)

The "doctors" in this team are powered by a new type of AI called Mamba.

The Analogy: Old AI models (like Transformers) are like a student trying to read a 1,000-page book by reading every single word against every other word to find connections. It's incredibly slow and uses a lot of energy.
The Mamba difference: Mamba is like a snake that slithers through the data. It remembers the important parts of the story as it moves forward without needing to re-read the whole book. It is incredibly efficient, allowing the system to run on smaller, cheaper computers (like the ones inside a prosthetic arm) without getting bogged down.

4. The "Voting Booth" (Classification)

Finally, the system doesn't just guess once. It breaks the muscle signal into tiny chunks and asks the team of experts to vote on what gesture is happening in each chunk.

The Analogy: If you ask 10 people what a blurry shape is, and 8 of them say "It's a cat," you trust the majority. MoEMba does this with "majority voting" to ignore the noisy, confusing parts of the signal and focus on the clear consensus.

The Result: Why is this a big deal?

In the past, these muscle-control systems worked great in a lab but failed in the real world because of that "day-to-day" variability.

The Old Way: If you trained the robot on Monday, it might not work on Tuesday because your skin was sweatier or you were more tired.
The MoEMba Way: Because it uses a team of specialists and understands the "music" of the muscles deeply, it stays accurate even when conditions change.

The Bottom Line:
MoEMba is a lightweight, efficient, and smart team of AI specialists that can listen to your muscles, ignore the daily noise, and correctly guess what you want to do—whether it's picking up a cup or playing a video game—even if you try it on a different day or with a different person. It's a major step toward making robotic arms and computer interfaces that actually work for real people in real life.

1. Problem Statement

The paper addresses the critical challenge of inter-session and inter-subject variability in High-Density Surface Electromyography (HD-sEMG) based hand gesture recognition.

The Core Issue: While Deep Learning (DL) models perform well in intra-session tasks (same subject, same recording session), their accuracy drops significantly (by 10–40%) when applied to new sessions or different subjects.
Causes of Variability: This degradation stems from non-stationary signal characteristics caused by muscle fatigue, skin impedance changes (sweating), electrode drift, anatomical differences, and limb positioning.
Limitations of Current Methods:
- Traditional ML: Relies on manual feature engineering (Time/Frequency domains) which struggles with the non-linear, multi-channel nature of sEMG.
- Deep Learning (CNNs/RNNs/Transformers): CNNs have limited receptive fields for long-range dependencies; RNNs struggle with long sequences; Transformers offer global context but suffer from quadratic computational complexity ( $O(n^2)$ ), making them inefficient for real-time, long-sequence sEMG data.
- Channel Independence: Many models treat sEMG channels as uncorrelated univariate time series, ignoring the crucial inter-channel dependencies (muscle coordination) inherent in gesture movements.

2. Methodology: The MoEMba Framework

The authors propose MoEMba, a novel framework combining Selective State-Space Models (SSMs) (specifically the Mamba architecture) with a Mixture of Experts (MoE) configuration and Wavelet Feature Modulation.

A. Architecture Overview

The framework consists of three primary modules:

Wavelet Transform Feature Modulation (WTFM) Block:
- Purpose: To capture multi-scale temporal and spatial relations while enhancing signal representation.
- Mechanism:
  - Processes signal patches using small ( $3\times3$ ) and large ( $7\times7$ ) temporal convolutional kernels to extract local and global features.
  - Applies Discrete Wavelet Transform (DWT) to local features, decomposing them into approximation and detail components (horizontal, vertical, diagonal).
  - Upsamples wavelet components and refines them using a Channel Attention Mechanism to focus on the most relevant channels.
  - Fuses these enhanced wavelet features with global features and raw channel embeddings to integrate both Time-Domain (TD) and Frequency-Domain (FD) information.
Mixture of Experts (MoE) Block:
- Purpose: To dynamically model diverse temporal patterns (short-term fluctuations and long-term trends) and adapt to distribution shifts.
- Mechanism:
  - Utilizes multiple Mamba experts (specialized SSM layers).
  - Employs a Sparse Gating Network to route input patches to the top- $k$ experts.
  - Optimization: Includes a tunable sparsity mechanism (adding Gaussian noise and selecting top- $k$ values) and a Balance Loss ( $L_B$ ) to prevent the gating network from biasing toward a single expert, ensuring uniform data distribution across experts.
  - Advantage: Mamba provides linear computational complexity ( $O(n)$ ) and global receptive fields, overcoming the quadratic cost of Transformers while maintaining the ability to model long sequences.
Classification Block:
- Processes the aggregated output from the MoE block.
- Uses a Majority Voting mechanism across signal patches belonging to the same gesture to determine the final class, reducing the impact of noisy predictions at the patch level.

3. Key Contributions

First Mamba Application for HD-sEMG: Introduces the first use of Mamba (an efficient SSM) for hand gesture recognition, leveraging its linear scaling for long time-series data.
Adaptive MoE Configuration: Proposes an adaptive combination of multiple Mamba experts to capture both short-term and long-term gesture dynamics, specifically addressing session-to-session variability.
Wavelet Feature Modulation (WTFM): Integrates wavelet decomposition with channel attention to capture multi-scale temporal and inter-channel spatial relations, enhancing the representation of non-stationary sEMG signals.
Computational Efficiency: The framework achieves high accuracy with significantly lower computational complexity (linear vs. quadratic) compared to Transformer-based SOTA models, making it suitable for resource-constrained embedded systems and real-time applications.

4. Experimental Results

Dataset: Evaluated on the CapgMyo HD-sEMG dataset (DB-b), which includes 10 subjects with recordings from two sessions (separated by $\ge$ 7 days) to test inter-session robustness.
Task: Classification of 8 distinct hand gestures.
Performance:
- Balanced Accuracy: MoEMba achieved 56.9% accuracy.
- Comparison: This significantly outperforms existing SOTA models (e.g., TVGGNet at 42.0%, NinaPro at 42.5%, and various Time-Domain methods ranging from 41.6% to 44.8%).
- Robustness: The model demonstrated superior consistency in handling the 10–40% variability typically seen between sessions.
Efficiency Metrics:
- Parameters: ~455,000.
- FLOPS: ~27.3 million (significantly lower than comparable Transformer models).
- Complexity: Linear $O(n)$ vs. Quadratic $O(n^2)$ for Transformers.
Class-Specific Performance: Achieved high AUC scores (e.g., 0.974 for "Thumb up"), though gestures with overlapping muscle activation patterns (e.g., partial finger flexions) showed slightly lower accuracy, which is expected given the signal complexity.

5. Significance and Conclusion

The MoEMba framework represents a significant advancement in Muscle-Computer Interfaces (MCIs) by solving the "session-to-session" generalization problem without requiring massive amounts of adaptation data or computationally expensive architectures.

Practical Impact: Its efficiency and robustness make it ideal for real-world applications such as active prosthetics, exoskeletons, and neurorehabilitation, where real-time processing and adaptability to changing user conditions are critical.
Future Directions: The authors suggest future work could involve using synthetic data to further augment training sets and integrating additional modalities to improve recognition reliability.

In summary, MoEMba successfully bridges the gap between high-accuracy deep learning and the computational/robustness constraints of practical HD-sEMG systems by leveraging the strengths of Mamba and Mixture of Experts.