Neural Networks and the Brain: What Does the Name Deceive Us About?

Do artificial neural networks work the way your brain does? The honest answer is that they borrowed an idea from the brain and then went down another road, much as an aeroplane resembles a sparrow in that both fly. Knowing the difference protects us from asking of them what they do not do well.

A borrowed idea, not a copy

It began in 1943, when Warren McCulloch and Walter Pitts offered a simplified mathematical model of the nerve cell, the first "artificial neuron". Then Frank Rosenblatt devised the perceptron in 1958, and the New York Times reported the US Navy as saying it was the "embryo" of an electronic computer expected to walk, talk and see. In 1969 Marvin Minsky and Seymour Papert showed that a single-layer perceptron cannot learn the exclusive-or (XOR) function, and interest and funding declined for years. It is often supposed that they refuted multi-layer networks, but they knew that such networks can do it. The lesson of the story is that exaggerated promises come back on their makers.

How wide is the gap?

According to a 2009 study by Azevedo and colleagues, the adult human brain contains about 86.1 ± 8.1 billion neurons (counting cells labelled with NeuN, in a sample of adult males). It makes up only about 2% of body weight but uses roughly 20% of its oxygen and calories, according to a 2002 review by Raichle and Gusnard.

But the number alone deceives. Beniaguev, Segev and London found that imitating the behaviour of a single cortical neuron (a computer model of a pyramidal cell from rat cortex, not a human cell) required a deep network of five to eight layers, and in most models around a thousand artificial nodes. In other words, a living cell computes more inside itself than a network node does. The name "neural" is a metaphor, not a description.

When do networks excel?

None of this belittles them. Demis Hassabis and John Jumper won the Nobel Prize in Chemistry in 2024 for AlphaFold2, which predicts protein structures, and David Baker shared it for designing new proteins. Utrecht University said they solved a "50-year-old problem". This is undoubtedly real capability.

Yet such achievements succeed where the goal is defined, the outcome measurable and the data plentiful. Here is the difference: a human generalises in uncontrolled settings and often learns from few examples, whereas networks usually need far more data and stumble outside what they were trained on. Szegedy and colleagues showed in 2013 that a network can be made to misclassify an image by a change the eye cannot perceive. That is a fragility human vision does not share in the same form.

Structural differences that change use

• Learning and operation: In most current models, learning stops at the end of training, and the network is then used as it is. The brain learns as it works.
• Understanding: What a network does is statistical optimisation of numerical weights, which produces intelligent behaviour without necessarily meaning understanding like ours. We addressed this confusion in the article "What Is Artificial Intelligence, Really? A Plain Explanation Without the Hype".

What is useful to us here?

With the right picture, judgement becomes easier. Use the network where it excels: specific, data-rich fields. Keep the human where context and responsibility are needed. Treat data as what the machine learns from, since it sees of the world only what is shown to it, and Arabic remains less present than other languages in much of that data, as discussed in the article "Arabic Artificial Intelligence: Do We Make or Consume? Sovereignty and Arabic Language Models".

The question "does the network imitate the brain?" overlooks the more useful one: what does this mathematical structure actually do, and where does it fall short? Whoever knows the tool's worth and limits uses it well; whoever takes it for a miniature mind entrusts it with what it should not be trusted with.