Fixing the problems of deep neural networks will require better training data and learning algorithms.

Area of Science:

• Computer Vision
• Neuroscience
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) are increasingly used to model biological vision.
• Concerns exist that DNNs may not accurately reflect human visual processing strategies.
• Previous work by Bowers et al. highlighted these discrepancies.

Purpose of the Study:

• To investigate the worsening problem of DNNs differing from human vision strategies.
• To propose methods for developing DNNs that reliably model biological vision.

Main Methods:

• Analysis of large-scale, high-accuracy DNNs.
• Comparison of DNNs' internal strategies with human visual processing.

Main Results:

• The discrepancy between DNNs and human vision strategies is increasing with DNN scale and accuracy.
• Current DNNs often achieve high accuracy using non-biological approaches.

Conclusions:

• DNNs, as currently developed, are inadequate models of biological vision.
• Methods are needed to guide DNN development towards biologically plausible solutions.
• Future research should focus on creating DNNs that emulate human visual strategies for better biological modeling.