Learning by passing the blame back

The perceptron could only learn with one layer of connections, and the 1969 critique had shown how limited that was. Networks with extra layers in the middle could in theory do much more, but nobody had a practical way to train them. If the network gave a wrong answer, which of its many inner connections was to blame?

In October 1986 the journal Nature published a short paper by David Rumelhart and Ronald Williams of the University of California, San Diego, and Geoffrey Hinton of Carnegie Mellon University. It was called "Learning representations by back-propagating errors".

The method works like this. The network makes a guess. You measure how far the guess was from the right answer. Then you pass that error backwards through the network, layer by layer, working out how much each connection contributed to it, and nudge every connection a little in the direction that would have made the answer better. Repeat this over many examples.

The paper showed that the hidden middle layers then learned useful features of the task on their own, without anyone telling them what to look for. Others had described similar maths before, but this paper made the case clearly and set off a wave of research.

Backpropagation, often shortened to backprop, is still how large neural networks learn, including the ones behind chatbots. At first it did not take over. The computers and data of the 1980s were too small, and it took more than twenty years before the method could show what it was capable of.

For a chatbot user, this is what "training" means. A chatbot was not given a rulebook. It made billions of guesses about text, was corrected each time, and its connections were adjusted bit by bit. That explains two everyday facts. It is good at whatever showed up often in its training. And it does not learn from your conversation in the moment: correcting it changes this chat, not the model, unless the company later uses that data to train a new version.

The idea behind this caseLesson: How machines learn →

1. 1Think of a chatbot's skills as habits built up from its training data, not as rules someone wrote.
2. 2Expect it to be strongest on common tasks and weakest on rare, local or very recent ones.
3. 3Do not assume that correcting a chatbot once has fixed it for good. Check whether your settings or the provider save or use your chats.
4. 4When it gets something wrong, give it the right information in the chat. That shapes this answer, even though it does not retrain the model.