Gradient boosted decision trees reveal nuances of auditory discrimination behavior

Area of Science:

• Animal behavior and machine learning
• Auditory perception in mammals
• Computational neuroscience

Background:

• Animal psychophysics generates large behavioral datasets.
• Gradient-boosted models offer insights into prediction mechanisms.
• Understanding decision-making in animal auditory tasks is crucial.

Purpose of the Study:

• To apply gradient-boosted models to ferret auditory psychophysics data.
• To investigate the impact of pitch variation on auditory stream segregation.
• To identify behavioral factors influencing ferret decision-making.

Main Methods:

• Training ferrets on a speech-based auditory task.
• Manipulating fundamental frequency (F0) to test pitch generalization and contribution to streaming.
• Implementing gradient-boosted regression and decision trees with SHAP analysis.
• Comparing model performance with traditional mixed-effects models.

Main Results:

• Ferrets performed accurately across pitch-shifted conditions.
• Within-trial pitch shifting increased false alarms and reaction times.
• Spectrotemporal similarity between target and non-target words predicted false alarms.
• Gradient-boosted models showed comparable or superior performance to mixed-effects models.

Conclusions:

• Gradient-boosted models provide valuable insights into animal decision-making.
• Pitch plays a subtle but significant role in auditory stream segregation.
• Spectrotemporal features are key predictors of errors in auditory tasks.
• Machine learning approaches enhance the analysis of complex behavioral data.