What are Neural Networks?

I want to push back on something before we even start. People hear 'neural network' and they assume it's a literal model of the brain. The lesson doesn't really disabuse them of that hard enough.

It uses the word 'inspired,' which is doing an enormous amount of lifting. Biological neurons are vastly more complex than anything in these diagrams. Calling it brain-inspired is roughly as accurate as calling a bicycle horse-inspired because both involve forward motion.

And yet the analogy is still useful. Inputs, weights, activation — there's genuine structural similarity even if the details don't hold.

I will concede that. The architecture borrows enough that the metaphor isn't useless. It just shouldn't be taken further than the architecture. The moment someone starts assuming neural networks therefore understand things the way brains do, the metaphor has done active damage.

Which is exactly the jump a lot of beginners make. They learn that weights adjust during training, they learn the word 'learn,' and suddenly the network has opinions.

It doesn't have opinions. It has numbers. Very many numbers, adjusted iteratively to reduce a loss function. The word 'learn' in this context is a convenient shorthand that has been catastrophically over-interpreted by approximately everyone who encountered it for the first time.

But the weights thing is actually the part worth dwelling on, because it's where the real insight is. The network isn't programmed with rules. It finds the rules, sort of, by being wrong repeatedly and correcting.

That is a reasonable way to put it. The weights begin essentially arbitrary, errors propagate backwards through the network, and adjustments accumulate across thousands or millions of iterations until the outputs become usefully accurate. It is, if you think about it, a rather brutal way to learn anything. Humans occasionally manage better.

Occasionally. The point about deep learning automating feature extraction is the one I think most people underestimate. Before deep learning, someone had to sit down and decide which features of an image mattered for identifying a cat. Now the network decides.

Which sounds like magic and is, in fact, matrix multiplication at scale. The features learned by early layers tend to be edges and textures. Later layers combine those into parts, and then into recognisable wholes. It's not mysterious — it's just happening at a speed and dimensionality that makes it feel mysterious.

Though I'd argue there's something genuinely interesting about the fact that we don't always know what features the middle layers have learned. That's not magic, but it's not nothing either.

That is a fair point and I won't pretend otherwise. Interpretability is a real and unsolved problem. The network has found something useful and we are, on occasion, not entirely sure what. I find that less enchanting and more concerning, but reasonable people disagree.

Which is probably where a beginner should land after this lesson — curious about that gap, not frightened by it and not over-romanticising it either.

If they come away understanding that weights encode learned relationships, that depth enables abstraction, and that none of this is cognition in any meaningful sense — that's sufficient. Most people manage one of those three. If you've managed all of them, you are, statistically, ahead of the curve.