Psychophysical Detection Testing with Bayesian Active Learning

Jacob R. Gardner, Xinyu Song, Kilian Q. Weinberger, Dennis L. Barbour, John P. CunninghamUAI 2015 · OpenReview

Abstract

Traditional psychophysical detection tests, such as standard hearing exams, are often inefficient, treating test points as independent and wasting time on redundant stimuli. This paper presents a novel Bayesian active learning framework that utilizes Gaussian Processes (GPs) to model latent sensory functions. By incorporating biological priors and a unique "OR-channel" likelihood for simultaneous multi-tone testing, the researchers demonstrate a significant reduction in the number of trials required—up to 85% in synthetic benchmarks—without sacrificing clinical accuracy.

Introduction: The Inefficiency of Grid-Based Testing

Psychophysical detection tests aim to map the "latent function" of human sensation, identifying the threshold at which a stimulus becomes perceivable. The conventional "grid search" approach, where a technician plays tones at fixed frequencies and intensities, suffers from three primary drawbacks:

  1. Inefficiency: Significant time is spent on stimuli that are clearly above or below the threshold.
  2. Information Silos: Each test point is treated independently, ignoring the inherent spatial correlations in sensory perception.
  3. Coarseness: Limited testing time results in a sparse sampling of frequencies, potentially overlooking specific sensory deficits.

The authors propose a more efficient, comfortable, and detailed alternative using active learning.

Methodology: Gaussian Processes and the "OR-Channel"

The core of the proposed method is a Gaussian Process (GP) model of the latent audiogram. A probit transformation is applied to the GP's continuous output to yield a probability of detection.

Incorporating Physical Priors

To improve efficiency, the model incorporates biological knowledge through its covariance function (kernel). A linear kernel is used for intensity to enforce the monotonic relationship that louder sounds are easier to detect. A squared exponential kernel is used for frequency, reflecting the smooth variation of hearing sensitivity across the spectrum.

Active Learning via Mutual Information

The algorithm selects the next stimulus by maximizing mutual information, effectively choosing the point that most reduces the model's uncertainty (entropy).

The "OR-Channel" Likelihood

A key innovation is the "OR-channel" likelihood, which allows for testing multiple stimuli simultaneously. If a subject hears any of the presented tones, they respond affirmatively. The authors demonstrate that querying multiple tones can be more informative than single tones for locating the perceptual boundary.

Information gain comparison: The "OR-channel" (paired tones, $I_2$) provides greater entropy reduction than single-tone queries ($I_1$) across the latent function $h(\cdot)$.

Evaluation and Key Results

The framework was validated through both synthetic benchmarking and real-world human testing.

Performance Gains

In synthetic trials, single-tone active learning required 50% fewer iterations than random sampling. The multi-tone (paired) strategy was even more effective, achieving an 85% reduction in iterations.

Clinical Validation

In human subjects, the GP model produced a high-quality approximation of the hearing threshold in as few as 15 iterations. The resulting 50% detection threshold was within 5dB of standard clinical audiograms, which typically require far more samples.

Comparison of a standard clinical audiogram (114 samples) with the multi-tone GP audiogram (60 iterations). The GP method captures a continuous probability of detection across the entire frequency-intensity space.

Discussion and Future Directions

While the results are promising, several areas for future work remain:

  • Computational Complexity: Selecting optimal sets for $k > 2$ tones is a combinatorial challenge that currently relies on greedy heuristics.
  • Human Factors: The cognitive load of responding to multiple simultaneous stimuli needs further investigation to ensure patient reliability.
  • Generalization: The specific kernels used were optimized for audition; adapting the framework to vision, pain, or touch would require defining appropriate sensory-specific priors.

By providing a full posterior distribution rather than a simple pass/fail threshold, this method offers clinicians a nuanced measure of certainty, paving the way for more precise and rapid sensory diagnostics.

Summary updated

BibTeX
@inproceedings{gardner2015psychophysical,
  author = {Jacob R. Gardner and Xinyu Song and Kilian Q. Weinberger and Dennis L. Barbour and John P. Cunningham},
  title = {{Psychophysical Detection Testing with Bayesian Active Learning}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2015},
  url = {https://openreview.net/forum?id=y7F1-o1JWr},
}
Figure