Which deep learning model can best explain object representations of within-category exemplars?

Area of Science:

• Cognitive Science
• Computer Vision
• Neuroscience

Background:

• Deep neural networks (DNNs) achieve human-level performance in object recognition tasks.
• Research has explored hierarchical similarities between human brain object representation and DNNs.
• The representational geometry of object exemplars within a single category in DNNs remains unclear.

Purpose of the Study:

• To investigate which DNN model best explains invariant within-category object representations.
• To compare the representational geometries of visual features in high-level DNN layers.
• To assess the invariability of within-category object representations by identifying object exemplars.

Main Methods:

• Computed similarity between representational geometries of visual features from high-level DNN layers.
• Utilized different DNN models, including transfer learning models based on ResNet50.
• Tested for invariability by identifying object exemplars within categories.

Main Results:

• Transfer learning models based on ResNet50 demonstrated the highest explanatory power for within-category object representation.
• ResNet50-based models also excelled in object identification tasks.
• The study identified specific object exemplars to test representational invariability.

Conclusions:

• The invariability of object representations in deep learning is more dependent on the quality of the transfer learning model than on the depth of the neural network.
• ResNet50-based transfer learning models provide a strong framework for understanding invariant object representations.
• Findings suggest a shift in focus towards optimizing transfer learning strategies for robust visual representations.