I was dreaming over the weekend (again) and I wondered… what if… what if datacenters were neural networks. Before you laugh and go somewhere else, let me explain why this idea would return the very best search results.
A neural network "learns" by being trained. There is a user who makes statements like: this is a human head. The neural network learns to recognise that image as a human head. The operator show thousands of different kinds of heads (and things that are not heads are shown as NOT a human head). Eventually the neural network begins to ask questions (is THIS a human head?) and the operator says yes or no. Over time the neural network (NN) can correctly identify a human head apart from a basket ball or fruit.
What if Google trained their datacenters to recognise good pages? Right now they are using inbound links to value a page, but that idea is time-limited. Google could use the user's clicks as training. The NN puts up a variety of pages with varying amounts of information and watches what people click on. With a cookie it can figure out which page I stayed on and which page I didn't. With this information (which Google already has) it can correlate the query (the search statement) to high performing pages.
Sure this would take time and money – but if the NN is able to correctly learn about what I am looking for and identify what page best meets my needs… then it's a killer search algo. The results page would be 100% accurate all of the time. It will always be learning how to serve up the best search results.