For human-like models, train on human-like tasks.

Area of Science:

• Cognitive Science
• Computer Vision
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) are increasingly used in vision research.
• Skepticism exists regarding DNNs' ability to model human vision accurately.
• Previous DNN models have failed to replicate psychological research findings.

Purpose of the Study:

• To propose a novel approach for evaluating deep neural networks (DNNs) as models of human vision.
• To investigate if training DNNs on human-like tasks enhances their ability to model human visual processing.
• To test the hypothesis that human-like training induces more human-like behaviors and representations in DNNs.

Main Methods:

• Training deep neural networks (DNNs) on tasks designed to mimic human visual perception challenges.
• Comparing the performance and internal representations of DNNs trained on human-like tasks versus standard tasks.
• Analyzing DNNs' ability to account for established psychological research findings in vision.

Main Results:

• Preliminary results suggest that human-like training can indeed elicit more human-like behaviors in DNNs.
• The internal representations within DNNs trained on human-like tasks show greater similarity to human visual representations.
• This approach provides a more robust framework for assessing the validity of DNNs as computational models of human vision.

Conclusions:

• Fairly assessing deep neural networks (DNNs) as models of human vision requires appropriate training paradigms.
• Training DNNs on human-like tasks is a promising strategy to improve their biological plausibility.
• Future research should focus on developing and implementing more sophisticated human-like training tasks for DNNs.