Approximating the Architecture of Visual Cortex in a Convolutional Network.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Neuroscience

Background:

• Deep convolutional neural networks (CNNs) share similarities with primate visual cortex but also exhibit significant differences.
• Reconciling these differences is crucial for developing more biologically plausible AI models.

Purpose of the Study:

• To develop a novel cortex-like CNN architecture.
• To ensure the architecture's consistency with empirical neural data.
• To create a framework for direct comparison between artificial and biological neural networks.

Main Methods:

• A specialized loss function was designed to quantify CNN architecture alignment with neural data (tract tracing, cell reconstruction, electrophysiology).
• Hyperparameter optimization was employed to minimize this loss function.
• Heuristics were developed for organizing network units into convolutional-layer grids.

Main Results:

• Optimized hyperparameters demonstrated consistency with neural data.
• The developed cortex-like CNN architecture features distinct characteristics, including longer skip connections, larger kernels and strides, and unique connection sparsity.
• Crucially, each layer in the cortex-like network corresponds directly to specific cortical neuron populations.

Conclusions:

• The developed cortex-like CNN architecture shows improved biological realism compared to typical CNNs.
• The one-to-one correspondence between network layers and cortical neuron populations enables precise future comparisons of model and brain representations.
• This work advances the development of more biologically realistic deep neural networks.