3 ms·

I think you don't recognize the power of an universal function approximator. You you seem to underestimate the possible number of ways that problems can be abst

I think you don't recognize the power of an universal function approximator. You you seem to underestimate the possible number of ways that problems can be abstracted as functions to and from n dimensional vectors. It's a very non intuitive statement that a neural network can be used to recognize dogs and images from images.

dchftcs 5y ago
The point is that there are many different universal function approximators, so just that adds no value to explaining why NNs are good. This type of theory is meaningful (because it's harder to have taken anything as complex as multi-layer perceptron seriously without it back in the 80s), but it's not a fact that's going to help us make any progress nowadays. Mentioning it over and over again just distracts us from the real problems we need to address.
YeGoblynQueenne 5y ago
Is it more intuitive to say that a convolutional neural network can map sets of pixels to scalars denoting categories of objects? I think not, but it's harder to misunderstand the less intuitive description, and much, much easier to overhype the more intuitive one. "OMG, neural nets can recognise cats and dogs!". Humans can recognise cats and dogs and also know what cats and dogs are, so it's very easy to misunderstand the ability of CNNs to map between sets of pixels and category labels as a general ability to understand the world, which is not there.