Structural differences between human and mouse neurons and their implementation in generative AIs.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computer Vision

Background:

• Brain function is dependent on neuronal network structures.
• Mouse and human brains exhibit distinct functional properties.
• Nanometer-scale 3D structures of mouse medial prefrontal cortex and human anterior cingulate cortex were analyzed.

Purpose of the Study:

• To compare mouse and human neuronal structures.
• To implement mouse neuron characteristics into artificial neural networks (ANNs).
• To evaluate the impact of biological neuron characteristics on AI image generation.

Main Methods:

• Comparative analysis of mouse and human brain tissue structures.
• Development of generative adversarial network (GAN) and denoising diffusion implicit model (DDIM) with mouse-mimetic parameters.
• Image generation tasks using diverse photo datasets (cat faces, cheese, human faces, birds, automobiles).
• Analysis of image entropy across datasets.

Main Results:

• Mouse neurons possess smaller somata and thinner neurites compared to human neurons.
• Mouse-mimetic GAN and DDIM showed improved performance on cat face and cheese datasets but underperformed on human face and bird datasets.
• Dataset image entropy influenced AI performance, suggesting a link between biological complexity and computational requirements.
• AI "preferences" aligned with common mouse associations.

Conclusions:

• The structural differences between mouse and human neurons impact ANN performance in image generation.
• Dataset characteristics, particularly image entropy, play a crucial role in AI model outcomes.
• Further research is needed to integrate more biological findings into ANNs to understand brain function and neuronal network relationships.