Inferring the basis of binaural detection with a modified autoencoder

Area of Science:

• Auditory Neuroscience
• Computational Audition
• Artificial Intelligence in Biology

Background:

• The human binaural system uses timing differences between ears to detect sounds in noise.
• Direct measurement of these binaural mechanisms in humans is challenging.
• Understanding these mechanisms is crucial for audiology and neuroscience.

Purpose of the Study:

• To investigate human binaural hearing mechanisms using interpretable artificial neural networks.
• To identify latent computational strategies employed by the brain for sound localization and detection.
• To explore the utility of explanatory AI models in auditory neuroscience.

Main Methods:

• Training modified autoencoder networks to perform a human-like binaural detection task.
• Analyzing the internal structure and dynamics of trained networks for interpretability.
• Comparing network-derived computations with known neural mechanisms in animal models.

Main Results:

• Optimal autoencoder networks developed artificial neurons sensitive to interaural timing cues.
• Network dynamics mirrored cross-correlation mechanisms, consistent with biological models.
• These emergent computations effectively explained human performance in binaural detection.

Conclusions:

• Interpretable neural networks can successfully infer complex auditory processing mechanisms.
• Cross-correlation-like computations are a general and effective solution for binaural signal detection.
• AI-driven models offer a powerful, non-invasive approach to studying human auditory perception.