What the hidden layer learns

Nobody tells the hidden layers what to look for. No engineer decides what matters. No expert writes down the intermediate steps. The hidden layers figure it out on their own, driven entirely by the training signal — by what the output needed, which came from what the examples required.

What emerges is often interpretable, in retrospect. In networks trained to recognize images, the first hidden layer tends to notice edges — differences in brightness between neighboring pixels. The next layer combines edges into shapes. The next combines shapes into parts of objects. By the final layer, the network is working with something that functions like the concept "dog" or "face" — though it would never use those words, and has no idea what they mean.

Nobody programmed those concepts. The network invented them because they were useful.

The knowledge isn't written down anywhere. It's discovered. That's the sharp line between this approach and the rule-based systems that came before.