Deep problems with neural network models of human vision.

Area of Science:

• Computational Neuroscience
• Cognitive Science
• Computer Vision

Background:

• Deep neural networks (DNNs) demonstrate high accuracy in image classification tasks.
• DNNs are frequently cited as superior models of biological vision based on predictive accuracy for human behavior and brain activity.
• Existing datasets often lack the experimental control to validate the specific features driving DNN predictions.

Purpose of the Study:

• To critically evaluate the claim that DNNs are the best models of human object recognition.
• To investigate whether DNNs' predictive success in vision tasks reflects biologically plausible mechanisms.
• To highlight the limitations of current DNN evaluation methods and propose alternative approaches for building better models of human vision.

Main Methods:

• Analysis of existing datasets used to compare DNNs with human performance in object recognition.
• Examination of DNN performance against psychological research findings, particularly those involving manipulated independent variables.
• Comparison of DNN-derived features with those relevant to biological vision and psychological theories.

Main Results:

• DNNs show limited overlap with biological vision mechanisms, despite high accuracy in image classification.
• DNNs account for a negligible proportion of results from psychological research on object recognition.
• Current evaluation metrics for DNNs in vision do not adequately test underlying hypotheses or biological plausibility.

Conclusions:

• The claim that DNNs are the best models of human object recognition is not supported by psychological data.
• Future research should prioritize developing models that explain psychological findings and test specific hypotheses through experimental manipulation.
• Promising alternative modeling approaches should focus on explaining psychological data to advance understanding of biological vision.