Emergence of Direction-Selective Retinal Cell Types in Task-Optimized Deep Learning Models.

Area of Science:

• Computational neuroscience
• Deep learning
• Retinal circuit modeling

Background:

• Convolutional neural networks (CNNs) excel at modeling sensory systems.
• Task-optimized CNNs explain retinal encoding and cell function.
• Understanding motion processing in the retina remains a challenge.

Purpose of the Study:

• To elucidate computational mechanisms underlying motion-selective retinal circuits using task-optimized CNNs.
• To develop and apply methods for reverse-engineering learned computational mechanisms in a biologically constrained CNN.

Main Methods:

• Designed a biologically constrained CNN for motion classification.
• Applied a toolbox of methods inspired by psychophysics, deep learning, and systems neuroscience.
• Reverse-engineered the model to identify emergent computational mechanisms.

Main Results:

• A computational mechanism emerged, involving direction-selective ganglion cells and starburst amacrine cells, for motion discrimination.
• These cell types, crucial for motion processing, were identified within the CNN.
• The model successfully discriminated between moving stimuli.

Conclusions:

• Direction-selective retinal circuits are ecologically designed for robust motion discrimination.
• Task-optimized CNNs offer a framework for interpretable AI and understanding neural computation.
• The study provides insights into the interplay between neural structure and function in motion perception.