The traditional way of building a neural network to learn a specific mapping from input vectors to output lables is to build a pyramid comprising layers of neurons. Each successive layer is notionally distilling the outputs of the previous one, funnelling the learnings towards the end answer - each layer being smaller than its inputs.
The new intuition is that it may be more effective to throw uniform sized learning layers on top of the inputs, like blankets piled on a bed, and giving a the assembled heap a significant learning period to digest the input in an undirected (labeless) manner. The final stage (to produce a ‘correct’ labeling) then consists of connecting up a single layer of ‘observer’ neurons that can pluck out the desired results from the mass of neurons that have smartened themselves up organically.
Feed-back / Feed-forward Symmetry
In a world where the information ‘understood’ by a network is captured by self-referential feed-forward and feed-backward iterations, the idea of imposing a top-down pyramid on the learning as a first step does not make as much sense as when designing for ‘traditional’ back-propagation.
The fact that we’re only interested in certain labels doesn’t mean that we can prune off all the other possibilities - doing this may be implicitly limiting the resolving power of our network.
blog comments powered by Disqus