Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Leveraging Recurrent Neural Networks for Predicting Motor Movements from Primate Motor Cortex Neural Recordings

Yuanxi Wang, Zuowen Wang, Shih-Chii Liu

TL;DR

An efficient deep learning solution for decoding motor movements from neural recordings in non-human primates using an Autoencoder Gated Recurrent Unit (AEGRU) model.

Abstract

This paper presents an efficient deep learning solution for decoding motor movements from neural recordings in non-human primates. An Autoencoder Gated Recurrent Unit (AEGRU) model was adopted as the model architecture for this task. The autoencoder is only used during the training stage to achieve better generalization. Together with the preprocessing techniques, our model achieved 0.71 $R^2$ score, surpassing the baseline models in Neurobench and is ranked first for $R^2$ in the IEEE BioCAS 2024 Grand Challenge on Neural Decoding. Model pruning is also applied leading to a reduction of 41.4% of the multiply-accumulate (MAC) operations with little change in the $R^2$ score compared to the unpruned model.

Leveraging Recurrent Neural Networks for Predicting Motor Movements from Primate Motor Cortex Neural Recordings

TL;DR

An efficient deep learning solution for decoding motor movements from neural recordings in non-human primates using an Autoencoder Gated Recurrent Unit (AEGRU) model.

Abstract

This paper presents an efficient deep learning solution for decoding motor movements from neural recordings in non-human primates. An Autoencoder Gated Recurrent Unit (AEGRU) model was adopted as the model architecture for this task. The autoencoder is only used during the training stage to achieve better generalization. Together with the preprocessing techniques, our model achieved 0.71 score, surpassing the baseline models in Neurobench and is ranked first for in the IEEE BioCAS 2024 Grand Challenge on Neural Decoding. Model pruning is also applied leading to a reduction of 41.4% of the multiply-accumulate (MAC) operations with little change in the score compared to the unpruned model.

Paper Structure

This paper contains 26 sections, 16 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Dataset
  4. Metrics
  5. $R^2$ accuracy
  6. Memory footprint
  7. Computational complexity
  8. Model architecture
  9. GRU
  10. Autoencoder
  11. AEGRU inference architecture
  12. AEGRU training
  13. Data preprocessing
  14. Neural activity representation
  15. Softplus and logarithm
  16. ...and 11 more sections

Figures (4)

  • Figure 1: Our AEGRU network architecture. Green blocks represent the inference pass, and the grey blocks compose the auxiliary branch used during training. The weight dimensions are indicated in the blocks. ‘bn'=batch normalization.
  • Figure 2: Data pre-processing pipeline. Each input represents $WS \times N \times 4$ ms of neural activity.
  • Figure 3: Heat map of mean $R^2$ on the test sets of the six recordings, evaluated across window sizes ($WS$) and number of steps ($N$).
  • Figure 4: Change in $R^2$ when the target pruning rate (TPR) is increased. The mean $R^2$ starts to decrease when TPR$> 0.6$.