In this article
Neural network vs human brain: the short answer is that an artificial neural network borrows one idea from the brain, that many simple units connected together can learn, and almost nothing else. The units, the way they learn, the energy they use and the way they remember are all different. Below we compare the two point by point, using published figures rather than slogans.
Key facts
- 86 billion neurons in an average adult human brain, plus about 85 billion other cells.
- About 20% of the body’s resting energy goes to the brain, roughly 20 watts.
- 175 billion parameters in GPT-3, a 2020 language model; its training used an estimated 1,287 MWh.
- Our calculation: that training energy would run a 20-watt brain for about 7,300 years.
Where the idea came from
In 1943 Warren McCulloch and Walter Pitts described a simplified “neuron” that adds up its inputs and fires if the total passes a threshold. In 1958 Frank Rosenblatt built the perceptron, a machine that learned its own weights. Today’s deep learning rests on a 1986 Nature paper by David Rumelhart, Geoffrey Hinton and Ronald Williams that popularised backpropagation, a way to adjust millions of weights at once. Each step took inspiration from neuroscience, but each also simplified it heavily. Our guide to what artificial intelligence is covers the wider history.
How an artificial neuron works
An artificial neuron is just arithmetic. It multiplies each input by a weight, adds the results, and passes the total through a simple function. A network stacks millions or billions of these in layers. “Learning” means nudging the weights until the outputs match examples. A parameter is one of those adjustable numbers.
A real neuron is a living cell. It receives chemical signals at thousands of connection points called synapses, and if enough arrive at once it sends an electrical spike down its axon. The timing of spikes carries information, chemicals called neuromodulators change how it responds, and its connections physically grow and shrink over time.
Neural network vs human brain: 7 differences
| Feature | Human brain | Artificial neural network |
|---|---|---|
| Basic unit | Neuron: a living cell with thousands of synapses | A weighted sum and a simple function |
| Scale | About 86 billion neurons | GPT-3: 175 billion parameters (weights, not neurons) |
| Signals | Electrical spikes timed in milliseconds | Continuous numbers passed layer by layer |
| Energy | About 20 watts at rest | One Nvidia H100 chip: up to 700 watts |
| Learning | Local changes at each synapse, continuously | Backpropagation of errors, mostly in a separate training phase |
| Memory | Stored in the same connections that compute | Weights in memory chips, moved to processors and back |
| After training | Keeps learning from single experiences | Usually frozen; new facts need retraining or extra tools |
Note that parameters are not neurons. A parameter is closer to the strength of one connection, so comparing 175 billion parameters with 86 billion neurons compares connections with cells. The brain has far more connections than neurons, because each neuron links to many others.
The energy gap, calculated
The Society for Neuroscience says the brain uses about 20% of the body’s energy at rest. With a resting adult using roughly 100 watts, that is about 20 watts, the power of a dim light bulb. By contrast, Nvidia lists its H100 data centre chip at up to 700 watts, and large models train on thousands of them.
A 2021 study led by Google researchers (Patterson et al.) reported that training GPT-3 used an estimated 1,287 megawatt-hours of electricity. Our calculation: 1,287,000 kilowatt-hours divided by 0.02 kilowatts is 64 million hours, or about 7,300 years of a brain running non-stop. A child learns language over a few years on that same 20 watts, although the comparison is rough, since the brain also had millions of years of evolution behind its starting wiring.
| Item | Power or energy | Source |
|---|---|---|
| Adult brain at rest | About 20 W | 20% of about 100 W resting use |
| Nvidia H100 (SXM) | Up to 700 W | Nvidia specifications |
| GPT-3 training run | 1,287 MWh | Patterson et al. (2021) |
| Brain-years equal to GPT-3 training | About 7,300 | World Elev calculation |
Where the analogy holds
- Learning from examples: both get better by adjusting connection strengths rather than following written rules.
- Layers of features: image networks build up edges, then shapes, then objects, which loosely resembles how the visual cortex processes images.
- Distributed knowledge: neither stores a fact in one place; it is spread across many connections.
- Graceful failure: damaging a few units or neurons usually degrades performance slightly rather than breaking everything.
Where it breaks
Backpropagation needs an exact error signal sent backwards through every layer, and most neuroscientists doubt the brain does this. Hinton himself has explored alternatives, such as the forward-forward algorithm, partly because backpropagation seems biologically unlikely. Brains also learn continually and from a single experience, while most networks are trained once and then frozen. Networks still fail at tasks people find easy, as our report on where AI still fails shows, and our explainer on what large language models can and cannot do covers why.
Frequently asked questions
How are neural networks similar to the brain?
Both are made of many simple connected units that learn by adjusting the strength of their connections, and both store knowledge spread across those connections.
What is the difference between the human brain and a neural network?
A neuron is a living cell that sends timed electrical spikes and keeps learning; an artificial neuron is a weighted sum. The brain also uses about 20 watts, far less than AI hardware.
Are neural networks inspired by the brain?
Yes. The first artificial neuron in 1943 and the perceptron in 1958 were directly modelled on simplified ideas about how neurons fire.
Is an AI neural network as big as the human brain?
Not in a like-for-like way. GPT-3 has 175 billion parameters, which are closer to connections than neurons, and the brain has far more connections than its 86 billion neurons.
Read more in our Artificial Intelligence section.
Sources
- Herculano-Houzel S. (2009): The human brain in numbers, Frontiers in Human Neuroscience
- BrainFacts (Society for Neuroscience): How much energy does the brain use?
- Brown et al. (2020): Language Models are Few-Shot Learners (GPT-3)
- Patterson et al. (2021): Carbon Emissions and Large Neural Network Training
- Nvidia: H100 Tensor Core GPU specifications
- Rumelhart, Hinton and Williams (1986): Learning representations by back-propagating errors, Nature
- Hinton G. (2022): The Forward-Forward Algorithm



