Abstract
A central goal of neuroscience is to understand how activity in the nervous system is related to features of the external world, or to features of the nervous system itself. A common approach is to model neural responses as a weighted combination of external features, or vice versa. The structure of the model weights can provide insight into neural representations. Often, neural input-output relationships are sparse, with only a few inputs contributing to the output. In part to account for such sparsity, structured regularizers are incorporated into model fitting optimization. However, by imposing priors, structured regularizers can make it difficult to interpret learned model parameters. Here, we investigate a simple, minimally structured model estimation method for accurate, unbiased estimation of sparse models based on Bootstrapped Adaptive Threshold Selection followed by ordinary least-squares refitting (BoATS). Through extensive numerical investigations, we show that this method often performs favorably compared to L1 and L2 regularizers. In particular, for a variety of model distributions and noise levels, BoATS more accurately recovers the parameters of sparse models, leading to more parsimonious explanations of outputs. Finally, we apply this method to the task of decoding human speech production from ECoG recordings.
Introduction: The Challenge of Sparsity in Neuroscience
In neuroscience, researchers frequently seek to map high-dimensional brain activity to external behaviors. These relationships are typically sparse, meaning that out of thousands of potential features (e.g., neurons or frequency bands), only a small subset is functionally relevant. To identify these features, standard statistical approaches employ structured regularizers like Lasso ($L_1$) or Ridge ($L_2$) regression.
While effective for prediction, these methods impose specific mathematical priors on the data. For example, Lasso assumes a Laplace prior, which forces many weights to zero but also "shrinks" the magnitudes of the remaining non-zero weights. This shrinkage introduces a systematic bias, making it difficult for scientists to interpret the true strength of neural connections.
The BoATS Framework
To address the limitations of traditional regularization, we propose Bootstrapped Adaptive Threshold Selection (BoATS). BoATS is a "minimally structured" method that separates the process of feature selection from the estimation of weight magnitudes. The algorithm proceeds in three primary stages:
- Null Distribution Estimation: The relationship between inputs and outputs is permuted (shuffled) to estimate a "noise floor" for the model weights. This identifies what weights would look like by chance.
- Adaptive Thresholding: An initial model is fit to the original data. Weights that do not significantly exceed the noise floor are set to zero. The optimal threshold is determined via cross-validation.
- OLS Refitting: The surviving features are refitted using Ordinary Least Squares (OLS). Because OLS lacks a regularization penalty, it provides unbiased estimates of the true weight magnitudes, avoiding the "shrinkage" characteristic of Lasso.
Performance and Parsimony
We evaluated BoATS against Ridge, Lasso, and Elastic Net using both synthetic simulations and real-world human electrocorticography (ECoG) data.
In synthetic tests, BoATS consistently outperformed traditional regularizers in recovering the true underlying model parameters. As shown in the comparative analysis, BoATS (referred to as ATS-OLS Refit in figures) achieves lower estimation error and superior model parsimony, as measured by the Bayesian Information Criterion (BIC). While Lasso often retains many small, noisy weights, BoATS effectively isolates the true signal.
Decoding Human Speech
The utility of BoATS was further demonstrated in decoding human speech production from ECoG recordings in the sensory-motor cortex. The task involved predicting the production of vowels (/a/, /i/, /u/) from high-frequency brain signals. BoATS-trained models successfully captured both across-vowel and within-vowel variability, reaching approximately 80% accuracy in predicting acoustic features. Crucially, the resulting models were more interpretable than those produced by Lasso, providing a clearer picture of the neural features driving speech production.
Discussion and Limitations
BoATS offers a robust alternative for sparse model estimation, particularly when unbiased parameter interpretation is critical. However, its performance can degrade in low-sample regimes where the number of observations is close to the number of features. In such cases, the strong priors of Lasso or Ridge may provide necessary stability. Future work will explore extending BoATS to non-linear modeling frameworks and broader applications in sensory neuroscience.
Summary updated