Computational evidence for an inverse relationship between retinal and brain complexity.

Area of Science:

• Computational neuroscience
• Visual neuroscience
• Machine learning

Background:

• An inverse relationship exists between brain and retinal complexity across species.
• Previous work suggested shallow networks encode complex features early, while deep networks encode simpler features.

Purpose of the Study:

• To extend computational explanations for the observed pattern of retinal adaptation.
• To analyze representational differences between shallow and deep neural networks in visual processing.

Main Methods:

• Representational analyses of neural network models.
• Comparison of feature encoding and decision boundaries in shallow versus deep networks.

Main Results:

• Shallow networks generate high-dimensional representations with linear decision boundaries, encoding specific visual features.
• Deep networks generate low-dimensional representations with nonlinear decision boundaries, encoding general visual features.
• Deep network representations require further processing for behavioral output.

Conclusions:

• Findings support the principle of simpler retinal features correlating with complex brains.
• A computational framework is offered for understanding neural network behavior in visual processing.