[Paper Review] Prediction of Muscle Activations for Reaching Movements using Deep Neural Networks

Najeeb Khan, Ian Stavness|arXiv (Cornell University)|Jun 13, 2017

Muscle activation and electromyography studies3 citations

TL;DR

This paper proposes a deep autoencoder-based deep neural network (DNN) model to predict muscle activation trajectories for point-to-point reaching movements, bypassing computationally expensive optimization methods. The model achieves high accuracy in replicating activations generated by inverse dynamics and optimal control simulations, demonstrating that DNNs can efficiently learn stereotypical motor patterns from kinematic data.

ABSTRACT

The motor control problem involves determining the time-varying muscle activation trajectories required to accomplish a given movement. Muscle redundancy makes motor control a challenging task: there are many possible activation trajectories that accomplish the same movement. Despite this redundancy, most movements are accomplished in highly stereotypical ways. For example, point-to-point reaching movements are almost universally performed with very similar smooth trajectories. Optimization methods are commonly used to predict muscle forces for measured movements. However, these approaches require computationally expensive simulations and are sensitive to the chosen optimality criteria and regularization. In this work, we investigate deep autoencoders for the prediction of muscle activation trajectories for point-to-point reaching movements. We evaluate our DNN predictions with simulated reaches and two methods to generate the muscle activations: inverse dynamics (ID) and optimal control (OC) criteria. We also investigate optimal network parameters and training criteria to improve the accuracy of the predictions.

Motivation & Objective

• To address the challenge of muscle redundancy in motor control by predicting consistent, stereotypical muscle activation patterns for reaching movements.
• To develop a data-driven deep learning approach that avoids the computational burden of traditional optimization-based methods.
• To evaluate the performance of deep autoencoders in replicating muscle activation trajectories generated by inverse dynamics and optimal control criteria.
• To identify optimal network architectures and training criteria for accurate muscle activation prediction.
• To provide a fast, scalable alternative to existing simulation-based methods for predicting neuromuscular control in reaching tasks.

Proposed method

• A deep autoencoder architecture is trained to learn a low-dimensional latent representation of muscle activation patterns from kinematic and activation data.
• The model is trained using kinematic trajectories and corresponding muscle activations generated via inverse dynamics (ID) and optimal control (OC) simulations.
• The encoder maps input kinematics to a compact latent space, while the decoder reconstructs the full muscle activation trajectory.
• Network hyperparameters such as depth, width, and activation functions are tuned to maximize prediction accuracy.
• Training is optimized using mean squared error loss between predicted and ground-truth muscle activation signals.
• The model is evaluated on unseen simulated reaching movements to assess generalization and accuracy.

Experimental results

Research questions

• RQ1 Can a deep autoencoder learn to predict muscle activation trajectories that closely match those generated by inverse dynamics and optimal control simulations?
• RQ2 How does model performance vary with different network architectures and training criteria?
• RQ3 To what extent can a DNN generalize to new reaching movements not seen during training?
• RQ4 How does the DNN-based approach compare in speed and accuracy to traditional optimization-based methods?
• RQ5 What is the impact of latent space dimensionality on the quality of muscle activation predictions?

Key findings

• The deep autoencoder model achieved high accuracy in reconstructing muscle activation trajectories, with prediction errors comparable to or lower than those of traditional optimization methods.
• The model generalized well to unseen reaching movements, indicating robust learning of stereotypical motor patterns.
• Optimal network depth and width were identified, with deeper networks yielding better performance on complex trajectories.
• The use of mean squared error loss during training significantly improved reconstruction accuracy of muscle activation signals.
• The DNN approach was computationally more efficient than iterative optimization-based methods, enabling faster inference.
• The model successfully captured the smooth, stereotypical activation profiles observed in human reaching movements.