Mapping the learning curves of deep learning networks.

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Machine Learning

Background:

• Interpreting deep neural networks (DNNs) is challenging, especially for non-tabular data.
• Current interpretation methods are often qualitative and lack systematic quantification.

Purpose of the Study:

• To introduce a cognitive science-inspired method for quantifying and visualizing DNN internal representations.
• To capture temporal dimensions of model learning: information-processing and developmental trajectories.

Main Methods:

• Developed a multi-dimensional quantification and visualization approach for DNN learning curves.
• Conducted 750 simulation runs on gesture detection and sentence classification tasks.
• Utilized four metrics (start, end-start, max, tmax) to quantify learning curves.

Main Results:

• Identified significant differences in learning patterns based on data sources and class distinctions (p < .0001).
• Revealed the role of spatial semantics in gesture learning and information gains in language learning.
• Highlighted non-monotonic progress, pairwise comparisons, and domain distinctions in learning curves.

Conclusions:

• The method provides insights into model appropriateness, input signal properties, and alignment with human learning.
• Offers a systematic way to analyze DNNs across different modalities and tasks.
• Has theoretical implications for understanding cognitive processing and multi-modal representations.