Training for object recognition with increasing spatial frequency: A comparison of deep learning with human vision.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Visual perception

Background:

• Human visual development and adult processing show a coarse-to-fine spatial frequency sensitivity.
• This progression, from low to high spatial frequencies, is crucial for interpreting visual input.

Purpose of the Study:

• To investigate if simulating a coarse-to-fine processing progression in deep convolutional neural networks (CNNs) impacts their visual representations.
• To compare CNN performance with this simulated progression against standard training and adult human visual representations.

Main Methods:

• Simulated coarse-to-fine processing in CNNs by gradually increasing spatial frequency information during training.
• Compared CNNs trained with standard and coarse-to-fine methods using behavioral and neuroimaging datasets.
• Evaluated CNN performance on low-pass-filtered, hybrid, and full spatial frequency images.

Main Results:

• Standard CNNs showed poor performance on low spatial frequency images, unlike humans.
• Coarse-to-fine training improved CNN classification accuracy from 0% to 32% on low-pass-filtered ImageNet images.
• Coarse-to-fine trained CNNs demonstrated increased sensitivity to low spatial frequencies in hybrid images.

Conclusions:

• The coarse-to-fine processing strategy is relevant for high-level object representations in artificial vision.
• Integrating computational, neural, and behavioral findings highlights the importance of spatial frequency variation processing.
• This approach enhances CNNs' ability to interpret visual information, mirroring aspects of human visual processing.