Neural Network Simulation of Infant Familiarization to Artificial Sentences: Rule-Like Behavior Without Explicit...

Area of Science:

• Cognitive Science
• Developmental Psychology
• Computational Neuroscience

Background:

• A central debate in cognitive science concerns whether symbolic rules or subsymbolic neural networks best model human cognitive processing.
• A prior study suggested infant language acquisition data uniquely supported symbolic rule learning.

Purpose of the Study:

• To demonstrate that unstructured neural networks can successfully simulate infant familiarization data from an artificial language.
• To challenge the conclusion that the data exclusively supports symbolic rule-based learning.

Main Methods:

• Simulated infant familiarization using unstructured neural networks with a sonority-based phoneme coding scheme.
• Compared network performance against existing infant data, focusing on attention, novelty preference, and generalization.

Main Results:

• Neural network simulations replicated key aspects of the infant data, including attention decreases, differential recovery to novel sentences, and extrapolative generalization.
• The models accounted for more data facets with fewer assumptions than previous accounts.
• The simulations showed that unstructured neural networks can exhibit behaviors previously attributed solely to symbolic rule learning.

Conclusions:

• The findings indicate that unstructured neural networks provide a viable alternative explanation for infant language acquisition data.
• The study supports the view that neural networks are capable of complex cognitive processing relevant to language development.
• The developed model offers predictions for future psychological research in infant cognition.