From biology to math
Where the idea comes from
"Neural network" is not a metaphor someone invented for marketing. In 1943, a neurophysiologist and a logician, Warren McCulloch and Walter Pitts, asked a genuinely strange question: could you describe what a brain cell does using nothing but arithmetic? Their answer became the blueprint every neural network still follows today.
A biological neuron does four things, in order:
- Dendrites receive electrical/chemical signals from other neurons
- The synapse at each connection can be weak or strong, that strength changes with experience
- The cell body (soma) adds up everything arriving at once
- If the combined signal is strong enough, the neuron fires a pulse down its axon to the next cells
McCulloch and Pitts replaced each of those four steps with a number or a simple operation. An artificial neuron does the exact same four things:
- Inputs: numbers instead of electrical signals (x₁, x₂, x₃...)
- Weights: a number per input, instead of synapse strength (w₁, w₂, w₃...)
- Weighted sum: multiply each input by its weight and add them up (Σ wᵢxᵢ)
- Activation, a function that decides whether (and how strongly) the neuron "fires" based on that sum
And that is the entire biological inspiration. Everything you'll learn in this module, layers, backpropagation, training, is this same four-step loop, repeated millions of times and tuned with math instead of chemistry.
Match the biology to the math
For each description of what happens inside a real neuron, pick which part of an artificial neuron plays the same role.
A dendrite picks up a signal arriving from a neighboring neuron.
A synapse has grown stronger after years of repeated use, the same input now produces a bigger effect than it used to.
Inside the cell body, every incoming signal is combined into a single running total.
The neuron stays silent for small signals, but fires a sharp electrical pulse the instant the combined input crosses a threshold.
The axon carries the pulse onward, to be picked up by the next neurons in line.
Two nearly-identical inputs arrive at two different synapses, and one produces a much larger internal effect than the other.
A weak, isolated signal on its own never makes the neuron fire, it only fires once enough of these weak signals add up together.
Where the analogy breaks down
It's tempting to stop here and think "okay, so an artificial neural network is basically a tiny digital brain." Resist that. The analogy is useful for building intuition about the shape of the idea, it is not an accurate description of what's actually running on a computer.
A few of the differences that actually matter:
- A biological neuron fires in continuous time, with precise timing that itself carries information, an artificial neuron just outputs one number per pass, no timing involved
- Your brain has roughly 86 billion neurons, each connected to thousands of others, and it keeps learning constantly, with no separate "training phase", most artificial networks have far fewer units and only update during an explicit training step
- Biological learning happens through complex chemistry (neurotransmitters, receptor changes), artificial "learning" is just adjusting numbers (the weights) using calculus, which you'll see in a later lesson ()
- A real neuron is a living, energy-consuming cell doing this as a side effect of being alive, an artificial neuron is a multiplication and an addition, nothing more
That last point isn't a shortcut the field settled for, it's an engineering trade-off. Computers are catastrophically bad at simulating chemistry: molecules, neurotransmitters, receptor binding, none of that runs fast on silicon. What computers are extraordinarily good at is one narrow thing: multiplying huge tables of numbers (matrices) as fast as possible, which is exactly what a GPU is built for. So instead of recreating biological chemistry, the field bent the math to fit the hardware. The neuron metaphor survived the translation; the chemistry didn't, because the hardware never gave it a choice.
So why keep the biological framing at all, if it's this loose? Because it answers the one question beginners actually get stuck on: why does a neural network look like that, with all those interconnected little units? The answer is historical, not mathematical, it's a structure borrowed from biology, then optimized ruthlessly for what math and hardware are good at. Once you know that, you can stop asking whether the network "thinks" like a brain, and start asking the more useful question: what is this weighted sum and activation function actually computing?
Check your understanding
A researcher says: "This weight is 0.9 instead of 0.1, so this connection matters more." Which biological property are they really describing?
What to remember
- ✓An artificial neuron is a four-step loop copied from biology: input → weight → weighted sum → activation → output.
- ✓Weight = synapse strength: a number that scales how much one input matters.
- ✓The weighted sum plays the role of the cell body, combining every input into one running total.
- ✓Activation plays the role of the firing threshold: it decides whether, and how strongly, the neuron responds.
- ✓The biological framing explains where the shape of the idea came from. It does not mean the network "thinks" like a brain, everything from here on is pure math.