Yueling Sun

Presenter: Yueling Sun

Title: Investigating Diagnostic Features in Prototype Category Learning Task: A Deep Learning Perspective

Abstract:

Category learning allows humans to classify and generalize novel stimuli. The prototype distortion task is commonly used to study this process, where category members are generated by distorting a prototype. While humans can categorize distorted stimuli, the diagnostic features they rely on are not well understood. To explore this, we investigated how a convolutional neural network (CNN) classifies artificial stimuli in this task. Using RUBubbles, a recently introduced prototype-based artificial dataset, we selected four prototypes from similarity space, generated distortions, and trained VGG16 for categorical learning. Results show that VGG16 achieved perfect classification and exhibited distinct activation patterns across the four learned categories.

After training, we applied the Gradient-weighted Class Activation Mapping (Grad-CAM) platform to identify the model’s diagnostic feature representations. By controlling heatmap thresholds, we tested the model’s performance by parametrically varying the amount of available information in these features. VGG16 was able to classify stimuli based solely on these features, demonstrating their key role in categorization. This feature utilization pattern offers a new perspective on diagnostic features important for human category learning, shedding light on how visual information is prioritized in prototype distortion tasks.