U-shaped learning and frequency effects in a multi-layered perceptron: implications for child language acquisition

Area of Science:

• Cognitive Science
• Computational Linguistics
• Artificial Intelligence

Background:

• The applicability of Parallel Distributed Processing (PDP) models to human cognition and language acquisition, specifically children's learning of inflectional morphology, is debated.
• Previous research has explored PDP models for tasks analogous to English verb stem and past tense mapping, with varying success.

Purpose of the Study:

• To investigate the capacity of a three-layer back-propagation network to learn various verb tense mappings (arbitrary, identity, vowel change, suffixation).
• To compare the performance of multi-layered perceptrons against single-layered perceptrons in this pattern association task.
• To explore how input conditions influence learning, error patterns, and the emergence of rule-like behavior and U-shaped learning.

Main Methods:

• Implemented a three-layer back-propagation network for pattern association tasks mimicking English verb tense morphology.
• Compared the performance of a multi-layered perceptron with a single-layered perceptron.
• Systematically varied input conditions, including mapping type, competition effects, verb stem frequency, and phonological subregularities.

Main Results:

• Multi-layered perceptrons demonstrated a superior capacity for learning verb tense mappings compared to single-layered perceptrons.
• Input characteristics significantly influenced network performance, affecting both correct and erroneous outputs.
• The study identified conditions under which rule-like behavior and U-shaped learning emerge in the network, mirroring child language acquisition.

Conclusions:

• Multi-layered back-propagation networks offer a viable model for understanding aspects of children's acquisition of English past tense morphology.
• Input properties play a crucial role in shaping learning trajectories and error patterns in artificial neural networks.
• The findings contribute to the ongoing discussion about the explanatory power of PDP models in cognitive and linguistic development.