Is a neural network like the brain?

The word “neural” is where most of the confusion starts. The name is honest about its origin: an artificial network is made of units that “loosely model the neurons in the brain.”^[1] The key word is loosely. A biological neuron is a living cell that does a great deal of its own processing. An artificial neuron multiplies some numbers, adds them up, and applies one function. The shared idea is only that a signal comes in, gets processed, and a signal goes out.

It is tempting to go further and treat a trained network as a small brain, but the research warns against that. An MIT study found that neural networks reproduced a known brain-like activity pattern only when researchers deliberately imposed assumptions that were, in their words, “inconsistent with the biology.” As one of the authors put it, “if the researchers hadn’t already known of the existence of grid cells, and guided the model to produce them, it would be very unlikely for them to appear as a natural consequence.”^[2]

Grid cells are real neurons, found in the brain’s entorhinal cortex, that “fire only when the animal is at one of the vertices of a triangular lattice,” letting a small number of cells encode a large number of positions for navigation.^[2] To get an artificial network to reproduce that pattern, earlier studies had to force each of its units to represent exactly one location, “which is not how biological place cells work: studies have shown that place cells in the hippocampus can respond to up to 20 different locations.”^[2] The match only appeared because the model was built to produce it, not because the network discovered how navigation actually works on its own.

So the analogy is useful for one thing: it explains why these systems are called neural networks. It is not evidence that a neural network thinks, understands, or works the way a brain does. Treat the name as a label for a math structure, and you will read the rest of this subject more clearly.