Cellular and subcellular specialization enables biology-constrained deep learning.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Synaptic plasticity underlies learning and memory, extensively studied at molecular and cellular levels.
• Artificial neural networks (ANNs) inform understanding of neural plasticity during learning, but their architectures and training algorithms lack biological compatibility.
• A gap exists in understanding how the brain coordinates learning across neural circuitry layers due to ANNs' limitations.

Purpose of the Study:

• To test a theory that biological learning relies on neuronal cell type specialization and compartmentalized dendritic signaling.
• To develop a biologically constrained artificial neural network (ANN) model for image classification.
• To bridge the gap between computational models and neuroscience principles for understanding brain learning.

Main Methods:

• Leveraging recent experimental evidence to inform ANN architecture and training.
• Developing a deep learning algorithm, dendritic target propagation, compatible with biological principles.
• Constructing multilayer ANNs with distinct excitatory and inhibitory cell types and compartmentalized neuronal units (soma and dendrites).

Main Results:

• Demonstrated accurate image classification using the biology-compatible deep learning algorithm.
• Showcased that ANNs with specialized cell types and compartmentalized units can learn effectively.
• The model adheres to strict biological constraints, enabling insights into neural learning mechanisms.

Conclusions:

• Biological learning may depend on the specialization of distinct neuronal cell types and compartmentalized dendritic signaling.
• Dendritic target propagation offers a biologically plausible mechanism for deep learning in neural circuits.
• The model provides testable predictions about neuronal cell type roles in coordinating learning across brain regions.