Chapter 5 of 25
Biological vs. Artificial Neurons
How much of the brain analogy is actually true?
Motivation
"Neural network" is a name that does a lot of marketing work. It conjures images of a silicon brain, and it's not entirely wrong to do so — the artificial neuron really was inspired by biological neurons. But the analogy is also routinely oversold, in ways that can leave you with a badly miscalibrated mental model right at the start of this course. Before you build a single artificial neuron, it's worth spending one chapter being precise about exactly which parts of the brain analogy hold up under scrutiny, and which parts quietly fall apart.
A Real-Life Scenario
Picture explaining to a curious friend how you can tell your dog's bark apart from the neighbor's dog's bark, even through a wall. Somewhere in your auditory system, signals from your ears travel through layers of neurons, each one deciding whether to pass a signal along based on input from thousands of others, shaped by a lifetime of hearing both dogs. Nobody designed that circuit. It emerged from experience, one adjusted connection at a time.
Now picture a voice-assistant app doing something superficially similar: distinguishing "Hey Siri" from background noise. Under the hood, it's running a network of artificial neurons, adjusting connection strengths during training so the right sound pattern produces the right response. The two systems share a common ancestor idea — signals combined and passed onward, connection strengths that change with experience — but they are built from utterly different substances, with utterly different rules. Understanding both halves of that sentence is the goal of this chapter.
Building the Intuition
The biological neuron
A biological neuron has four parts worth knowing by name, because their artificial counterparts will follow you through the rest of this course.
Dendrites
Branch-like structures that receive electrical signals from other neurons.
Cell body (Soma)
Integrates — sums up — the incoming signals.
Axon
A long fiber that transmits the neuron's output signal onward, if the integrated signal is strong enough.
Synapses
Connection points between one neuron's axon and another's dendrites, where connection strength can change through learning.
A biological neuron "fires" — sends a signal down its axon — only when the combined input arriving at its dendrites exceeds a certain threshold. This all-or-nothing firing behavior is precisely what inspired the threshold-based activation of the very earliest artificial neuron models, back in the McCulloch-Pitts neuron you met in the previous chapter.
The artificial neuron
The artificial neuron takes this same shape and reduces it to arithmetic: multiply, sum, and apply a nonlinearity.
Where: are the input values, are learnable weights (the analogue of synaptic strength), is a learnable bias (loosely, the analogue of a firing threshold, with the sign convention flipped), is the pre-activation weighted sum, is a nonlinear activation function — you'll meet several concrete choices for in a later chapter — and is the neuron's output.
Intuition: The weighted sum lets the neuron learn which inputs matter more (large ) and which matter less (small ), while the nonlinearity determines how strongly the neuron "fires" in response to that combined signal.
Mapping the Analogy
Put the two side by side, and the inspiration is obvious:
| Biological Concept | Role | Artificial Analogue |
|---|---|---|
| Dendrites | Receive signals from other neurons | Input values $x_i$ |
| Synaptic strength | Determines how strongly a signal is transmitted | Weight $w_i$ |
| Cell body (soma) | Integrates incoming signals | Weighted sum $z = \sum w_i x_i + b$ |
| Firing threshold | Determines whether the neuron fires | Activation function $\varphi(z)$ |
| Axon | Transmits output to other neurons | Output value $y$, passed to the next layer |
| Learning (synaptic plasticity) | Connection strengths change with experience | Weight updates via gradient descent / backpropagation |
This table is genuinely useful as a first mental model, and it's why the name "neural network" stuck in the first place. But treat it as a starting sketch, not a blueprint — the next section is about where the sketch stops matching the territory.
Where the Analogy Breaks Down
It's tempting — and extremely common in popular writing — to describe artificial neural networks as literal simulations of the brain. That framing is misleading. The artificial neuron is a drastic mathematical simplification, and modern deep learning owes far more to statistics, linear algebra, and calculus than to neuroscience. Knowing the differences below matters for interviews as much as for building accurate intuition.
| Aspect | Biological Neural Network | Artificial Neural Network |
|---|---|---|
| Signal type | Discrete electrical spikes (action potentials); timing matters | Continuous real-valued numbers; timing (in a standard MLP) doesn't matter |
| Learning rule | Not fully understood; complex biochemical processes like long-term potentiation | A precisely defined mathematical procedure: gradient descent via backpropagation |
| Number of neurons | Human brain: roughly 86 billion neurons | Modern MLPs: typically thousands to a few million units |
| Connectivity | Highly recurrent, sparse, and irregular | Typically feedforward and fully connected between layers (in a vanilla MLP) |
| Energy efficiency | Extremely efficient — roughly 20 watts for the entire brain | Training large networks can consume enormous amounts of energy |
| Global objective | No single, explicit global loss function being minimized | An explicit, well-defined loss function minimized via gradient-based optimization |
Every one of these differences is worth sitting with for a moment, because each one quietly explains a design choice you'll see later in this course. The fact that artificial learning is a precisely defined mathematical procedure (row two) is exactly why backpropagation, covered soon, can be derived and proven correct — nothing so precise exists yet for biological learning. The fact that a standard MLP is feedforward and fully connected (row four) is a deliberate engineering simplification, not a claim about how the brain is wired.
"Are artificial neural networks accurate models of the brain?" The correct, nuanced answer: they are loosely inspired by biological neurons — particularly the ideas of weighted connections and threshold-like firing — but they are a major simplification. Modern deep learning is best understood as an application of mathematical optimization, built from calculus and linear algebra, rather than as a branch of neuroscience.
Key Takeaways
- Biological neurons integrate signals arriving at their dendrites and fire down the axon once that integrated signal crosses a threshold. - The artificial neuron mirrors this structure with a weighted sum, a bias term, and a nonlinear activation function. - Weights are analogous to synaptic strength; the activation function is loosely analogous to the firing threshold. - The analogy is a genuinely useful starting intuition, but breaks down in important ways: artificial networks use continuous values, a precisely defined mathematical learning rule (gradient descent), and typically far simpler, feedforward connectivity compared to the brain's massively recurrent wiring. - Deep learning is best understood as applied mathematical optimization, not as neuroscience — the brain inspired the shape of the neuron, not the mechanics of how it learns.
Common Mistakes
Describing a neural network as "just like a brain" glosses over enormous differences in scale, signal type, connectivity, and — most importantly — the learning mechanism itself. Use the analogy to build initial intuition, then let the precise math (starting in the next chapter) take over.
Modern MLPs, even large ones, use orders of magnitude fewer units than the human brain, wired in a far simpler pattern. Scale alone doesn't close that structural gap.
Interview Corner
Where This Shows Up in Practice
Every time you hear a neural network described as "mimicking the brain" in a news article or product pitch, you now have the tools to evaluate that claim precisely: yes, in the loose structural sense of weighted connections and threshold-like firing; no, in the deeper sense of learning mechanism, connectivity, scale, or energy efficiency. That precision matters practically, too — it's exactly why the rest of this course builds the Multi-Layer Perceptron from calculus and linear algebra rather than from neuroscience, and why the coming chapters on weights, activations, and backpropagation will feel like engineering and mathematics, not biology.