Do neural networks think like brains?

No, and the comparison is worth taking apart rather than dismissing, because the places where it fails are more interesting than the places where it holds. The word "neural" is a historical accident that has been doing public-relations work for seventy years.

Where the resemblance came from

The original 1943 model was an explicit simplification of a neuron: inputs arrive, get weighted, and if the total passes a threshold the unit fires. That was a reasonable sketch of what was known about neurons at the time, and it is roughly where the resemblance stops.

What real neurons do that artificial ones do not

• Spike in time. Real neurons communicate with discrete pulses whose timing carries information. Almost all artificial units output a continuous number and have no concept of when.
• Chemistry. Neuromodulators change how whole regions behave on timescales from milliseconds to hours. There is no equivalent of a mood in a standard network.
• Structural change. Brains grow and lose connections throughout life. Artificial networks almost always have fixed wiring and only change the strengths.
• Local learning. Backpropagation, which trains nearly every network you have heard of, requires each connection to know an error computed globally and sent backwards. Real synapses have no access to that, and nobody has found the brain doing anything closely equivalent.
• Scale and sparsity. A human brain has in the order of 10¹¹ neurons with 10¹⁴ connections, mostly local. Large artificial networks are far smaller in units and far denser in connections.

What does hold up

Three things, and they are not nothing. Both systems compute with many simple units in parallel rather than one complicated sequential one. Both store what they know in connection strengths rather than in addressable memory. Both degrade gracefully: damage a few units and performance drops a little rather than failing outright.

Those are real structural parallels, and they are why the networks work at all. They are not evidence that the mechanisms are the same.

The version in this project, which is closer in one specific way

The organisms here are unusual among artificial networks in that nobody trains them by backpropagation. Their weights are inherited, mutated and filtered by whether their carriers leave descendants - which is, mechanically, how brains came to be the way they are.

Some of them also change during their own lifetime, by a local rule that only uses information available at the connection itself. That is far closer to what a synapse could plausibly implement than backpropagation is.

So this world gets closer on how the weights came to be, and remains just as far away on spikes, chemistry and structural growth. A fair summary: evolution and local learning are the two parts of the brain story that artificial networks usually skip, and this project runs both while skipping everything else.

What a neural network actually is · Evolution against learning, measured separately · Open one and look