Semi-supervised learning of a nonnative phonetic contrast: How much feedback is enough?

Area of Science:

• Cognitive Psychology
• Machine Learning
• Neuroscience

Background:

• Semi-supervised learning is common in machine learning due to reduced labeled data needs.
• Human semi-supervised learning is less understood, with few documented instances.
• This study investigates human semi-supervised learning in a non-native phonetic classification task.

Purpose of the Study:

• To investigate human semi-supervised learning mechanisms.
• To determine the optimal ratio of feedback to no-feedback trials for learning.
• To explore the role of no-feedback trials in learning outcomes.

Main Methods:

• Participants performed a non-native phonetic classification task over two days.
• Training involved varying numbers of feedback and no-feedback trials daily.
• Control conditions included stimulus exposure without feedback.

Main Results:

• Performance did not improve with 60 feedback trials daily alone.
• Performance significantly improved when 60 feedback trials were paired with 240 no-feedback trials daily.
• Increasing feedback trials to 240 did not further enhance learning; decreasing to 30 abolished it.
• No-feedback testing, not mere stimulus exposure, was crucial for learning.

Conclusions:

• Human semi-supervised learning is feasible and influenced by the structure of training.
• A specific ratio of feedback to no-feedback trials is critical for effective learning.
• No-feedback testing periods may act as a trigger for incorporating information from non-feedback trials into learning.