The Biological Inspiration
Every field of engineering borrows from nature. Aeroplanes borrow from birds. Sonar borrows from bats. Artificial neural networks borrow from the most complex structure we know of in the universe: the human brain.
Understanding this borrowing (what was taken directly, what was simplified and what was left behind) gives you a mental model for how artificial networks work that no amount of matrix algebra can replace.
The Biological Neuron
Your brain contains roughly 86 billion neurons. Each one is a specialised cell that receives signals, processes them and either fires a signal of its own or stays quiet. The basic anatomy of a neuron has three parts.
Dendrites are the branching input arms of the neuron. They receive electrochemical signals from other neurons across tiny gaps called synapses. A single neuron can have thousands of dendritic connections.
The cell body (soma) accumulates the incoming signals. If the combined signal exceeds a threshold (a tipping point), the neuron fires. If it does not reach that threshold, nothing happens. This is an all-or-nothing response.
The axon is the output wire. When the neuron fires, it sends an electrical pulse (an action potential) down the axon. The axon branches at the end, connecting via synapses to the dendrites of other neurons.
Synapses and Signal Strength
Not all connections are equal. Some synaptic connections are strong: a small signal arriving there significantly increases the chance that the receiving neuron will fire. Other connections are weak: a large signal barely makes a difference.
This variation in synaptic strength is how the brain encodes information. When you learn something, the strengths of specific synaptic connections change. This process is called synaptic plasticity and the neuroscientist Donald Hebb described it in 1949 with a phrase that has become a mantra in machine learning: "neurons that fire together, wire together." Connections that are used repeatedly become stronger.
How the Brain Processes Signals
A sensory experience: say, touching something hot: sends a cascade of signals through your nervous system. Sensory neurons fire, passing signals to interneurons in your spinal cord and brain, which in turn pass signals to motor neurons that cause your hand to pull away. This all happens in milliseconds, long before you are consciously aware of pain.
Higher-level processing (recognising a face, understanding a sentence, solving a maths problem) involves coordinated activity across billions of neurons in multiple brain regions simultaneously. The brain is massively parallel: it does not process things one step at a time the way a traditional computer does.
The Leap to Artificial Neurons
In 1943, neurophysiologist Warren McCulloch and mathematician Walter Pitts published a paper that asked: could the basic logic of a neuron be implemented mathematically?
They proposed a simple model. An artificial neuron receives several numerical inputs. Each input is multiplied by a weight, which represents the strength of that connection. The weighted inputs are summed. If the sum exceeds a threshold, the neuron outputs a 1 (fires). If not, it outputs a 0 (stays silent).
This is the McCulloch-Pitts neuron and it is the ancestor of every modern neural network.
What Was Preserved and What Was Simplified
The biological analogy is powerful but imperfect.
Preserved: the idea of weighted inputs, the threshold for firing, the layered organisation of neurons, the modification of connection strengths through learning.
Simplified away: the complex electrochemical dynamics of real neurons, the time-dependent nature of biological signals, the physical geometry of the brain, the enormous diversity of neuron types, the role of glial cells and much more.
Modern artificial neural networks are not simulations of the brain. They are mathematical abstractions inspired by the brain's general logic. This distinction matters. Artificial networks have capabilities the brain does not (perfect recall, parallelisable training) and lack capabilities the brain has (common sense, general reasoning, learning from one example).
The metaphor is useful. But do not let it mislead you into thinking we understand the brain by building neural networks or that neural networks work the way brains do.
Quiz: What is synaptic plasticity and how does it relate to the concept of weights in an artificial neural network?