IR80 wrote:
But they must be interpreting/summarising based on something programmed?
So, my very, very basic lay person understanding is that machine learning doesn’t require programming, but instead a training set.
So when we were asked to prove we were not a bot when accessing a website, by selecting all the squares containing a road sign, we were helping to create a training set for a self-driving car. Presumably there are now algorithms that can recognise road signs, thanks to our efforts, so that security feature for weeding out bots has become redundant.
The training sets are used to test automatically generated, pretty much random algorithms. Millions of them. And most of them are rubbish. But if you pick the best 1%... well, they’ll still be rubbish. But if you use them as the basis to generate millions more and pick the best 1% of them, survival of the fittest style, and so on, and so on... you get there.
So I guess the human input is in creating the environment and providing the nutrients (the training set), but the algorithm evolves more than being programmed. For the latest game machine learning for Chess and Go, I think the machine created its own training set. Rather than being fed loads of games from top human players, and learning to mimic them super efficiently, it started from scratch playing against itself.
