Dendrites: Why Biological Neurons Are Deep Neural Networks

Biological neurons—especially their dendrites—are far more than simple “wires” that sum inputs. Voltage-gated ion channels and dendritic nonlinearities let single neurons perform computations long thought to require multi-layer artificial neural networks, including time-sensitive pattern processing and even XOR-like logic. That matters because it reframes how brains implement learning and inference: computation can happen inside one cell, not just across networks of many neurons.

The discussion starts by contrasting early machine-learning neuron models with real cellular physiology. A perceptron resembles a neuron’s output mechanism: voltage-gated sodium channels can create an all-or-none action potential once membrane voltage crosses a threshold, followed by potassium channels that restore the resting state. But the perceptron’s weakness is input handling. Traditional textbook descriptions treat dendrites as passive, leaky cables that attenuate signals and effectively weight synaptic inputs by distance and receptor strength. That picture breaks down once dendrites are recognized as electrically active structures packed with voltage-gated channels.

Dendrites contain sodium channels that support backpropagation—action-potential-like activity traveling from the axon region back into dendritic branches—helping drive synaptic plasticity. They can also generate local, small depolarizations that transiently amplify synaptic inputs. A key coincidence detector is the NMDA receptor: it opens only when both neurotransmitter is present and the membrane is sufficiently depolarized. Because NMDA channels allow calcium influx (along with sodium), they produce NMDA spikes on longer timescales (hundreds of milliseconds) and enable nonlinear integration.

Those nonlinearities give dendrites computational reach. Dendritic processing can discriminate the order of incoming spikes: sequential activation in one direction can yield a different electrical and chemical response than activation in reverse. The system is also sensitive to activation velocity, enabling sequence-selective outputs. Evidence is cited that NMDA-driven dynamics can enhance stimulus selectivity in the visual cortex of awake animals, linking dendritic computations to behaviorally relevant processing.

A highlight comes from a 2020 study led by Matthew Larkum, which reported dendritic calcium action potentials in human layer 2/3 pyramidal neurons. These calcium spikes appear only within a narrow “just right” range of excitatory input strength—too weak fails to reach calcium-channel activation thresholds, too strong prevents spikes—creating a built-in selectivity that supports logic-like operations. Using two synaptic input groups (A and B), the neuron can respond when either group alone triggers a dendritic spike, but not when both are activated together—an XOR pattern.

Finally, the transcript connects these biophysical mechanisms to deep learning. A separate paper, “Single cortical neurons as deep artificial neural networks,” builds a detailed biophysical neuron model and trains deep convolutional neural networks to predict its output spikes from synaptic inputs. The learned network needs roughly 5–8 layers when NMDA channels are included, but collapses to about one hidden layer when NMDA channels are removed—directly tying dendritic nonlinearities to the computational depth required. The convolutional model also generalizes to synaptic configurations it never saw, suggesting it can infer underlying biophysics from training data. The overall takeaway: single neurons can act like deep, nonlinear computational units, making dendrites central to how brains compute.