| Sumario: | Research on the kinds of features that a neural network is likely to find useful when making identifications is discussed. Neural networks are computer programs that are based loosely on animals' nervous systems, and as such, they are able to learn from experience. At the moment, however, the only way to teach them is by supervised learning, whereby the computer is shown both a set of data and a label telling it what the picture represents. It would be easier, and probably profitable, to be able to do the equivalent of telling a neural network to recognize data from an obvious feature. Geoff Hinton of University College London, in England, thinks he has come up with a way to do this. Rather than asking a person to choose the features, he is asking another neural network. They way in which Hinton's neural network operates, how it is different from other neural networks, the fact that it replicates the human failing of seeing what it expects to see regardless of the truth, and how this might, in itself, be valuable knowledge about neural networks is discussed.
|
|---|