A machine learning approach to identify stochastic resonance in human perceptual thresholds

Abstract

Background: Stochastic resonance (SR) is achieved when a faint signal is improved with the addition of the appropriate amount of white noise. Perceptual thresholds are expected to follow a characteristic performance improvement curve as a function of the white noise level added (i.e., thresholds are reduced with an optimal amount of added white noise, beyond which excessive white noise is no longer beneficial). Since SR exhibition in perceptual thresholds is defined by a shape rather than a statistical difference, the presence of SR is typically identified qualitatively. The current state-of-the-art is for blinded human judges to categorize the presence of SR by visually examining data. While categorizations are made with subject data intermixed within a balanced, simulated dataset, which accounts for false positives, this method is still subjective and prone to human error.

New method: We use a logistic regression (LR) trained on engineered features in order to quantitatively classify exhibition of SR. The LR was trained on datasets simulated from a model for SR performance enhancement.

Results: We implemented the algorithmic classification process in 6 perceptual threshold test cases, informed by the literature and parameters were defined by experimental subject data. Comparison to Existing Method(s): We report algorithmic classifications of SR exhibition, considering the 6 test cases, that outperform existing subjective methods in accuracy (p < 0.05).

Conclusions: We demonstrate that algorithmic classification can effectively identify SR in perceptual thresholds, providing a rigorous, objective, and quantitative approach to identifying the presence of SR.

Keywords: Classification; False positives; Logistic regression; Perceptual thresholds; Stochastic resonance (SR); White noise.