When forgetting fosters learning: A neural network model for statistical learning.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Psycholinguistics

Background:

• Learning continuous signals requires segmenting them into discrete units, like words in speech.
• Statistical learning, particularly transitional probabilities (TPs), is a key mechanism for identifying these units.
• Human statistical learning exhibits flexibility, including sensitivity to forward/backward TPs and novel units.

Purpose of the Study:

• To explain the sophisticated computational abilities observed in human statistical learning.
• To model how transitional probabilities (TPs) are computed and utilized for unit segmentation.
• To investigate the role of forgetting in enabling flexible statistical learning.

Main Methods:

• A computational model with tunable Hebbian excitatory connections and inhibitory interactions was developed.
• The model's learning dynamics were analyzed under varying degrees of forgetting (weak, intermediate, strong).
• Model performance was evaluated against known hallmarks of human statistical learning.

Main Results:

• Intermediate levels of forgetting were essential for the model to replicate human-like TP computations.
• The model demonstrated sensitivity to both forward and backward TPs, and learned novel units.
• Sophisticated statistical learning was achieved even without a dedicated memory store for item encoding.

Conclusions:

• Forgetting is a critical factor determining the flexibility and sophistication of statistical learning.
• The proposed model provides a parsimonious explanation for key features of statistical learning.
• Statistical learning mechanisms may operate independently of explicit memory encoding processes.