Distribution-free prediction intervals with conformal prediction for acoustical estimation

Area of Science:

• Signal Processing
• Acoustics
• Machine Learning

Background:

• Acoustical parameter estimation is crucial across various fields.
• Existing methods lack confidence measures for estimates, hindering real-world application.
• External uncertainties like noise and sensor errors impact estimation performance.

Purpose of the Study:

• To introduce a method for statistically valid uncertainty quantification in acoustical parameter estimation.
• To adapt conformal prediction (CP) for models lacking training data or using analytical approaches.
• To address the challenge of limited calibration data distribution in CP.

Main Methods:

• Utilizing conformal prediction (CP) for uncertainty interval generation.
• Applying CP to both data-driven and analytical acoustical models.
• Validating performance on direction-of-arrival and source localization tasks.

Main Results:

• Conformal prediction (CP) successfully generated statistically valid uncertainty intervals.
• Demonstrated CP's applicability to acoustical parameter estimation with various uncertainty sources (noise, interference, sensor location).
• Showcased CP integration with both data-driven and traditional propagation models.

Conclusions:

• Conformal prediction (CP) offers a robust solution for uncertainty quantification in acoustical parameter estimation.
• CP provides statistically valid confidence intervals essential for reliable deployment.
• The method is effective even when calibration data differs from test-time data distributions.