How Computationally Complex Is a Single Neuron? | Quanta Magazine

computational-neuroscienceneural-networksdeep-learningdendrites

Abstraction: Single biological neuron equivalent to 5–8 layer deep neural network

Key points:

Researchers at Hebrew University (Beniaguev, Segev, London) trained a deep neural network to mimic a simulated pyramidal neuron and found it required 5–8 layers and ~1,000 artificial neurons per biological neuron
The complexity came primarily from dendritic trees and NMDA receptors on dendrite surfaces — findings consistent with prior 2003 work showing dendrites compute as a two-layer network
Result implies loose analogy of "one artificial neuron = one biological neuron" is wrong; a 50-layer image classification network may correspond to only ~10 real neurons
Timothy Lillicrap of DeepMind noted the result forces a rethinking of how artificial and biological neurons are compared
The authors propose replacing each simple unit in deep networks with a 5-layer "mini network" modeled on a neuron, though computational advantage remains unproven
Real neurons may be even more complex — current neuroscience cannot record full input-output functions of live neurons

Connections: Deepmind · Hebrew University Of Jerusalem · Computational Neuroscience · Neural Networks · Deep Learning