Biological vs. Artificial Neurons

Structure

A biological neuron has three main parts. Dendrites are branching tree-like structures that receive input signals from other neurons through synapses. The cell body (soma) integrates these inputs. The axon transmits the output signal (action potential) to other neurons, potentially over distances of more than a meter in the spinal cord.

A single biological neuron can have 1,000 to 10,000 synaptic inputs, each with its own strength, type (excitatory or inhibitory), and temporal dynamics. The dendritic tree is not passive wiring; it performs local computation. Individual dendritic branches can implement AND and OR logic, detect coincidences between inputs, and generate local calcium spikes that amplify signals. A single biological neuron is computationally more powerful than a single artificial neuron, more closely resembling a small neural network.

An artificial neuron has no physical structure. It is a mathematical function: output = f(sum(wi * xi) + b). It receives a vector of input values, multiplies each by a weight, sums the results, adds a bias, and applies an activation function. The entire computation happens instantaneously in a single operation. There are no dendrites, no axon, no spatial arrangement.

Communication

Biological neurons communicate through action potentials (spikes), all-or-nothing electrical pulses that travel along the axon at speeds of 1 to 100 meters per second. Information is encoded in the timing and frequency of these spikes, not in continuous values. A neuron firing at 100 spikes per second represents a stronger signal than one firing at 10 spikes per second. The precise timing of individual spikes relative to other neurons' spikes may carry additional information.

At the synapse, an arriving spike triggers the release of neurotransmitter molecules that cross the synaptic cleft and bind to receptors on the receiving neuron. Different neurotransmitters have different effects: glutamate excites, GABA inhibits, dopamine modulates learning, serotonin affects mood. A single biological neuron may release and respond to multiple neurotransmitter types, each with different dynamics.

Artificial neurons communicate by passing continuous floating-point values. The output of one neuron becomes the input of connected neurons in the next layer, with no temporal dynamics, no neurotransmitters, and no distinction between excitatory and inhibitory pathways (positive and negative weights serve this role implicitly). Communication is instantaneous and deterministic.

Learning Mechanisms

Biological learning involves changes in synaptic strength (synaptic plasticity) through multiple mechanisms operating at different timescales. Short-term plasticity changes synapse strength for milliseconds to minutes. Long-term potentiation (LTP) strengthens synapses for hours to years when pre-synaptic and post-synaptic neurons are co-active. Long-term depression (LTD) weakens synapses under different activity patterns. Structural plasticity creates new synapses or eliminates existing ones.

These changes are governed by local rules: each synapse adjusts based on the activity of its two connected neurons and the local concentration of neuromodulators. There is no global error signal that propagates backward through the network. Whether something analogous to backpropagation occurs in the brain is an active debate in neuroscience, but the known plasticity mechanisms are fundamentally local.

Artificial learning uses backpropagation and gradient descent, which require a global error signal that propagates from the output through every layer to every weight. This is computationally powerful but biologically implausible: the brain has no known mechanism for transmitting precise gradient information backward through multiple synaptic layers.

Energy Consumption

A biological neuron consumes roughly 10 to 100 nanowatts. The brain's 86 billion neurons collectively use about 20 watts. This extreme efficiency comes from event-driven computation (neurons only consume energy when they spike), analog signal processing (continuous membrane potential changes use less energy than digital floating-point operations), and co-located memory and computation (synaptic weights are stored at the same location where computation occurs, eliminating the energy cost of moving data).

An artificial neuron's energy consumption depends on the hardware. On a modern GPU, a single multiply-accumulate operation (the core of an artificial neuron's computation) consumes roughly 1 picojoule. But the total energy cost includes memory access (moving weight values from memory to the processor), which typically dominates at 100-1,000 picojoules per access. The von Neumann bottleneck, the separation of memory and computation, is the primary reason digital neural networks are far less energy-efficient than biological ones.

Speed

Individual biological neurons are slow. A spike takes about 1 millisecond, and maximum firing rates are 200-500 spikes per second. The brain compensates with massive parallelism: billions of neurons operate simultaneously.

Artificial neurons on GPUs compute in nanoseconds, roughly a million times faster than biological neurons. A GPU can perform trillions of multiply-accumulate operations per second. However, for real-time processing of sensory data, the brain matches or exceeds digital systems because its parallelism (86 billion neurons) far exceeds GPU parallelism (thousands of cores), and its event-driven processing avoids redundant computation on unchanged inputs.

What This Comparison Teaches Us

The comparison reveals that artificial neural networks capture the abstract principle of biological computation (interconnected units that learn by adjusting connection strengths) while discarding most of the biological details (electrochemistry, temporal dynamics, neuromodulation, structural plasticity). This abstraction has been enormously successful for practical applications, but the discarded details likely encode important computational principles that AI has not yet discovered.

The brain's energy efficiency, continual learning ability, sample efficiency (learning from few examples), and robustness to damage all exceed what artificial neural networks achieve. Whether reproducing the discarded biological mechanisms would close these gaps, or whether entirely different engineering solutions are needed, remains an open question driving both neuroscience and AI research.