Computational Modeling of Ganglion Cell Bicolor Opponent Receptive Fields and FPGA Adaptation for Parallel Arrays

Area of Science:

• Computational neuroscience
• Biomimetic computing
• Visual system modeling

Background:

• Biological systems exhibit remarkable low power consumption and high parallelism.
• Mimicking these biological efficiencies in artificial systems is challenging, especially when ignoring lower-level information processing.
• Understanding visual pathways from the retina to the lateral geniculate nucleus (LGN) is key to developing efficient artificial vision.

Purpose of the Study:

• To model the K, M, and P visual pathways at a fine-grained level.
• To achieve efficient information transmission with minimized energy consumption.
• To implement a circuit-level distributed parallel computing model for visual processing.

Main Methods:

• Modeling the visual system from the retina to the LGN, focusing on K, M, and P pathways.
• Developing a fine-grained computational model for efficient information processing.
• Implementing the model on Field-Programmable Gate Arrays (FPGAs) using a distributed parallel computing approach.

Main Results:

• Successful transfer of visual information with low energy consumption and high parallelism.
• Achieved maximum frequency of 200 MHz and parallelism of 600 on Artix-7 FPGAs.
• Demonstrated a power consumption of only 0.142 W per single receptive field model.

Conclusions:

• The developed model effectively replicates biological visual processing efficiencies.
• The low power and high parallelism make it suitable for resource-constrained assistive vision devices.
• This approach offers a pathway for creating compact and efficient artificial vision systems.