Combining convolutional neural networks and cognitive models to predict novel object recognition in humans

Area of Science:

• Cognitive Psychology
• Computer Vision
• Computational Neuroscience

Background:

• Convolutional neural networks (CNNs) excel at object recognition in natural images.
• Previous research utilized novel objects like Greebles, Ziggerins, and Sheinbugs in recognition studies.

Purpose of the Study:

• To determine if CNN object representations can predict human performance on novel objects.
• To integrate CNN representations with a cognitive decision-making model (LBA) for predicting errors and response times.

Main Methods:

• Utilized object representations from CNNs trained on natural images.
• Drove a linear ballistic accumulator (LBA) cognitive model with CNN representations.
• Analyzed performance in a same-different matching task involving novel objects.

Main Results:

• A combined CNN-LBA model reasonably accounted for human errors and response times.
• Model predictions were accurate at both average and individual participant/item levels.
• High-level CNN representations were effective in predicting recognition performance.

Conclusions:

• CNN representations show promise for understanding visual object recognition.
• Integrating CNNs with cognitive models can explain complex human behaviors.
• This approach extends cognitive modeling capabilities for visual tasks.