Learning mechanisms in cue reweighting.

Area of Science:

• Psycholinguistics
• Computational Neuroscience
• Speech Perception

Background:

• Feedback effectively shifts attention to perceptual cues in speech perception.
• The underlying learning mechanisms for this attention shift remain unclear.
• Existing models include supervised error-driven learning and reinforcement learning.

Purpose of the Study:

• To compare supervised error-driven learning and reinforcement learning predictions for feedback-driven cue weighting in speech perception.
• To investigate the role of alternative cues and cue predictiveness in learning.
• To determine the computational principles guiding attention shifts in phonetic learning.

Main Methods:

• Computational simulations comparing supervised and reinforcement learning models.
• Behavioral experiments with native English listeners exposed to bimodal/unimodal VOT distributions.
• Manipulating the predictiveness of Voice Onset Time (VOT) and the availability of an alternative cue (F0) for voicing.

Main Results:

• Reinforcement learning accurately predicts observed cue weighting changes.
• Supervised learning predictions were not supported by the experimental data.
• Listeners reduced reliance on VOT only when an alternative cue (F0) could predict voicing.
• Frequency distributions of cues did not influence learning outcomes.

Conclusions:

• Reinforcement learning principles, not supervised error-driven learning, explain feedback's effect on cue weighting in speech perception.
• Attention shifting in learning phonetic cues for phonological categories is driven by reinforcement learning.
• These findings highlight the role of alternative cue availability and positive evidence in learning.