Deep Pinsker and James-Stein Neural Networks for Decoding Motor Intentions From Limited Data

Marko Angjelichinoski, Mohammadreza Soltani, John Choi, Bijan Pesaran, Vahid TarokhIEEE Transactions on Neural Systems and Rehabilitation Engineering 2021 · DOI

Abstract

Non-parametric regression has been shown to be useful in extracting relevant features from Local Field Potential (LFP) signals for decoding motor intentions. Yet, in many instances, brain-computer interfaces (BCIs) rely on simple classification methods, circumventing deep neural networks (DNNs) due to limited training data. This paper leverages the robustness of several important results in non-parametric regression to harness the potentials of deep learning in limited data setups. We consider a solution that combines Pinsker's theorem as well as its adaptively optimal counterpart due to James-Stein for feature extraction from LFPs, followed by a DNN for classifying motor intentions. We apply our approach to the problem of decoding eye movement intentions from LFPs collected in macaque cortex while the animals perform memory-guided visual saccades to one of eight target locations. The results demonstrate that a DNN classifier trained over the Pinsker features outperforms the benchmark method based on linear discriminant analysis (LDA) trained over the same features.

Introduction

Deep Neural Networks (DNNs) are the gold standard for many classification tasks, but they face significant hurdles in neural engineering due to their high data requirements. Collecting high-quality brain data, specifically Local Field Potentials (LFPs) via invasive electrodes, is time-consuming, expensive, and technically challenging.

Furthermore, brain signals are inherently non-stationary, with statistical properties that shift over time and across different cortical areas. When training data is scarce, complex DNNs are prone to overfitting, often leading researchers to rely on simpler, less capable tools like Linear Discriminant Analysis (LDA). This research addresses a critical gap: how to harness the power of deep learning for motor intention decoding when limited to small, noisy datasets.

Methods

The researchers propose two novel architectures: the Deep Pinsker Neural Network (DPNN) and the Deep James-Stein Neural Network (DJSNN). The core innovation lies in a mathematically-grounded feature extraction stage that reduces the learning burden on the neural network:

  1. Mathematical Denoising: Instead of requiring a network to learn noise filtration from scratch, the model utilizes established theorems from non-parametric regression, treating the LFP signal as a smooth function embedded in Gaussian noise.
  2. Fourier Projection: Raw LFP signals are projected onto a Fourier basis (sines and cosines).
  3. Optimal Shrinkage:
    • Pinsker’s Theorem provides a method to optimally "shrink" or weight Fourier coefficients to minimize error, acting as a mathematically optimized low-pass filter.
  4. James-Stein Estimation offers an adaptive version that does not require prior knowledge of the signal's smoothness, increasing flexibility for real-world applications.
  5. Classification: These refined, low-dimensional features are fed into a relatively shallow Multi-Layer Perceptron (MLP) for final classification.

By using rigorous mathematics for feature extraction, the DNN only needs to learn the final mapping to motor intentions, significantly reducing the number of trainable parameters.

Experimental Setup and Datasets

The study utilized LFP data from two macaque monkeys (Monkey A and Monkey S) performing a memory-guided visual saccade task. Data was collected using a 32-channel microelectrode array implanted in the prefrontal cortex (PFC).

Experimental task sequence: A) Start, B) Target On, C) Target Off, D) Memory Cue, E) Saccade movement, F) End.

As individual recording sessions often provided too few trials for deep learning, the authors employed a "data bundling" algorithm. This grouped trials from similar electrode depths to create robust training sets of approximately 1,400 trials each.

Key Insights

  • Phase Information: Unlike traditional methods that focus solely on wave amplitude (power), Pinsker-based features preserve phase information. This additional layer of data is a primary driver for performance gains over standard spectrogram-based methods.
  • Efficiency and Speed: The DPNN achieved results comparable to EEGNet, a popular CNN for brain signals. However, because the DPNN uses fixed mathematical filters rather than trainable convolutional layers, it trains faster and is less prone to overfitting on small datasets.
  • Frequency Relevance: The study confirmed that the most relevant information for decoding eye movements is stored in frequencies below 10 Hz. The Pinsker approach naturally emphasizes these frequencies while suppressing high-frequency noise.

Results

  • Superior Accuracy: The DPNN consistently outperformed the benchmark Pinsker-LDA method across all cortical depths. In Monkey A, the DPNN reached 91.8% accuracy at superficial sites, compared to 84.6% for the LDA version.
  • Robustness at Depth: The performance gap between the DNN-based approach and linear methods widened as electrodes moved deeper into the brain (where signal-to-noise ratios typically drop), suggesting that non-linear deep learning is better at identifying patterns in noisier data.
  • Computational Edge: Compared to EEGNet, the DPNN maintained a slight accuracy advantage while significantly reducing the number of hyperparameters requiring tuning, making it a more practical solution for BCI deployment.

Conclusion and Future Work

While the data bundling technique proved effective, it assumes signal similarity across nearby cortical sites, which may not always hold in highly specialized regions. Future research will focus on cross-subject generalization, incorporating "trainable" Pinsker weights to allow the network to fine-tune mathematical filters, and applying these extractors to non-invasive modalities like EEG.

Summary updated

Abstract

Non-parametric regression has been shown to be useful in extracting relevant features from Local Field Potential (LFP) signals for decoding motor intentions. Yet, in many instances, brain-computer interfaces (BCIs) rely on simple classification methods, circumventing deep neural networks (DNNs) due to limited training data. This paper leverages the robustness of several important results in non-parametric regression to harness the potentials of deep learning in limited data setups. We consider a solution that combines Pinsker's theorem as well as its adaptively optimal counterpart due to James-Stein for feature extraction from LFPs, followed by a DNN for classifying motor intentions. We apply our approach to the problem of decoding eye movement intentions from LFPs collected in macaque cortex while the animals perform memory-guided visual saccades to one of eight target locations. The results demonstrate that a DNN classifier trained over the Pinsker features outperforms the benchmark method based on linear discriminant analysis (LDA) trained over the same features.
BibTeX
@article{angjelichinoski2021deep,
  author = {Marko Angjelichinoski and Mohammadreza Soltani and John Choi and Bijan Pesaran and Vahid Tarokh},
  title = {{Deep Pinsker and James-Stein Neural Networks for Decoding Motor Intentions From Limited Data}},
  journal = {IEEE Transactions on Neural Systems and Rehabilitation Engineering},
  year = {2021},
  doi = {10.1109/tnsre.2021.3083755},
}
Figure