Bottom-up skill learning

This thesis proposes a unified, domain-general account of early language acquisition using the PRIMs cognitive architecture. While traditional theories are often divided into learning-based and rule-based perspectives, this research suggests that infants seamlessly integrate these processes through bottom-up learning.

Instead of predefined, complex rules, the model uses primitive processing elements that self-organize into processing sequences. Simulations show that the same architecture can produce both lexical and syntactic behaviors depending on the input patterns and processing efficiency.

The thesis moves away from rigid "if-then" rules toward procedural representations. Learning occurs through associations between real-time model contexts and processing elements. Additionally, to bridge the gap between habit and goal-oriented processing, the framework introduces skill elements.

These light-weight abstractions allow the model to differentiate task-specific outcomes and adapt to situational changes. The simulation studies predict that infants begin with macro-level processing (continuous encoding) before fine-tuning into specific patterns. It also accounts for preferential looking behaviors by linking gaze duration to the efficiency and complexity of cognitive processing. Finally, the research suggests that cognitive load influences the developmental onsets for different linguistic tasks.