MouseNet: A biologically constrained convolutional neural network model for the mouse visual cortex.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Systems neuroscience

Background:

• Convolutional neural networks (CNNs) model mammalian visual systems, but typically primate ventral stream architecture.
• Mouse and primate visual systems differ in hierarchical depth, raising questions about architectural roles in visual computation.

Purpose of the Study:

• To introduce a novel framework for a biologically constrained CNN model of the mouse visual cortex.
• To investigate the computational capabilities and neural representations of a mouse-visual-cortex-inspired network.

Main Methods:

• Constructed a CNN (MouseNet) using experimental data: interareal connectome, neuron counts, and layer connection statistics.
• Evaluated MouseNet on an image classification task (ImageNet) and compared its performance to VGG16.
• Utilized representational similarity analysis with the Allen Brain Observatory Visual Coding dataset to compare MouseNet's representations to mouse visual cortex activity.

Main Results:

• MouseNet achieved approximately 2/3rds the performance of VGG16 on ImageNet, demonstrating computational viability.
• Representational similarity analysis showed MouseNet recapitulates neural representations in the mouse visual cortex.
• Task training improved physiological quantity distributions closer to observed mouse brain data, and further optimization did not significantly increase biological similarity.

Conclusions:

• Biologically constrained models like MouseNet offer a viable framework for studying mouse visual cortex computation.
• Architectural constraints derived from the mouse brain can support effective visual processing.
• Task optimization alone may not be sufficient to fully capture biological neural representations.