What it could not do

Now imagine a clue that only means something in relation to another clue.

The suspect bought a train ticket the morning of the murder. On its own, that means nothing. People buy train tickets. Score it low.

But the murder happened on that train.

Suddenly the ticket is everything. The first clue didn't change. The second clue changed what the first one meant.

The detective's method breaks here. There's no way to score "bought a train ticket" correctly in isolation. The meaning only exists in the combination. Neither fact alone contains it.

The Perceptron is stuck in the same way. It scores each input on its own and adds them up. It has no way to ask what two inputs mean together, or what one tells you about the other.

When the answer lives in a relationship between inputs rather than in any single input, the Perceptron can't find it. Not because it needs more training. Because the method itself can't look for that kind of thing.