Pyramidal neuron as 2-layer neural network

Bartlett Mel, Department of Biomedical Engineering, University of Southern California, USA

The pyramidal neuron is the principal cell type in the mammalian forebrain, but its function remains poorly understood. Using a detailed compartmental model of a hippocampal CA1 pyramidal cell calibrated with a broad spectrum of in vitro data, we recorded responses to stimuli consisting of up to 100 high-frequency-activated excitatory and inhibitory synapses distributed across the apical dendritic tree.  Surprisingly, we found the cell's mean firing rate could be predicted by a simple formula that happens to also describe a conventional 2-layer "neural network''. In the first layer, synaptic inputs drive several dozen separately thresholded sigmoidal subunits---physically corresponding to the long, thin terminal dendrites that make up the bulk of the cell's receptive surface. In the second layer, subunit outputs are summed within the main trunk and cell body prior to final thresholding. We found the 2-layer model explains 91% of the variance in our spike rate data, compared to 37% for a "point neuron'' model. We conclude that a 2-layer network with sigmoidal hidden units may provide a useful abstraction for the computing function of an individual pyramidal neuron. We end by considering some of the implications of the two-layer model for the information processing functions of cortical tissue.

Joint work with Panayiota Poirazi and Terrence Brannon