Exploring the effectiveness of reward-based learning strategies for second-language speech sounds

Area of Science:

• Cognitive Science
• Neuroscience
• Computational Linguistics

Background:

• Adults often struggle with non-native speech category learning in traditional settings.
• A video game paradigm shows efficient learning when sounds have functional significance.
• Behavioral and neural data suggest reinforcement learning mechanisms are involved in speech category acquisition.

Purpose of the Study:

• To computationally formalize and test the hypothesis that reinforcement learning underlies speech category learning.
• To compare a deep reinforcement learning network with a supervised model for speech learning.
• To investigate the role of specific neural circuitry in effective speech sound learning.

Main Methods:

• Implemented a deep reinforcement learning network to model the mapping between environmental input and actions.
• Compared the reinforcement network's performance to a supervised learning model.
• Evaluated models using two experiments: learning synthesized auditory noise tokens and improving speech sound discrimination.

Main Results:

• The reinforcement network's behavior closely matched human performance in both experiments.
• Both reinforcement and supervised models showed comparable performance.
• The similarity in model outputs suggests no inherent computational advantage for reward-based learning alone.

Conclusions:

• Reinforcement learning mechanisms are implicated in adult speech category learning.
• The specific neural circuitry, particularly links between the striatum and superior temporal areas, is crucial for effective learning.
• While reinforcement learning models human behavior, the computational benefit over supervised learning may be minimal, highlighting the importance of neural mechanisms.