Single cortical neurons as deep artificial neural networks

Area of Science:

• Computational neuroscience
• Machine learning applications
• Neural network modeling

Background:

• Understanding the computational complexity of single neurons is crucial for deciphering brain function.
• Biophysical models offer detailed neuronal dynamics but can be computationally intensive to analyze.
• Machine learning, particularly deep neural networks (DNNs), presents a powerful tool for complex system analysis.

Purpose of the Study:

• To develop a systematic machine learning approach for characterizing the input/output (I/O) mapping complexity of neurons.
• To quantify the computational resources required to model neuronal I/O functions using DNNs.
• To investigate how specific biophysical properties, like NMDA receptor function, influence neuronal computational complexity.

Main Methods:

• Training deep neural networks (DNNs), including temporally convolutional and fully connected architectures, to replicate the I/O functions of biophysical neuron models.
• Utilizing millisecond resolution data to capture spiking dynamics.
• Analyzing the trained DNNs' weight matrices to interpret the underlying computational mechanisms.

Main Results:

• A multi-layered temporally convolutional DNN was necessary to accurately model a layer 5 cortical pyramidal cell (L5PC), demonstrating significant I/O mapping complexity.
• The trained DNN exhibited robust generalization to input data outside the training distribution.
• Removing NMDA receptors drastically simplified the required network architecture, indicating their substantial contribution to computational complexity.
• Analysis of DNN weights suggested that synaptic integration in dendrites can be viewed as spatiotemporal pattern matching.

Conclusions:

• DNNs provide a powerful framework for unifying the characterization of single-neuron computational complexity.
• The complexity of neuronal I/O mapping is highly dependent on specific biophysical properties.
• Cortical neurons possess unique computational architectures that may underpin the brain's powerful information processing capabilities.