Schizophrenia-Mimicking Layers Outperform Conventional Neural Network Layers

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Psychiatry

Background:

• Nanometer-scale 3D studies reveal schizophrenia patients have thin, tortuous neurites compared to controls.
• This suggests suppressed connections between distal neurons in brain microcircuits.

Purpose of the Study:

• To design a schizophrenia-mimicking artificial neural network (ANN) based on observed biological alterations.
• To simulate and evaluate the functional impact of these connection alterations in ANN models.

Main Methods:

• Developed a novel "schizophrenia connection layer" replacing fully connected layers in ANNs.
• Tested the network's performance on image classification tasks using MNIST and CIFAR-10 datasets.
• Investigated the role of weight matrix shape (band matrices) and applied a "schizophrenia convolution layer" in VGG configurations.

Main Results:

• The schizophrenia connection layer showed tolerance to overfitting and outperformed standard fully connected layers.
• Performance enhancement was linked to the use of band matrices as weight windows.
• A schizophrenia convolution layer allowed for significant weight elimination (60%) in VGG networks without accuracy loss.

Conclusions:

• Schizophrenia-associated neural connection alterations can be functionally beneficial in ANNs.
• These findings suggest that observed brain network changes in schizophrenia may play a role in cognitive function.
• Schizophrenia layers offer a practical alternative to conventional layers, easily integrated into existing network architectures.