| title | layout |
|---|---|
Creating deep-learning networks |
default |
While single-threaded and multithreaded deep-learning networks are just two ways of distributing the task load, real differences arise between single network and multinetwork deep-learning systems, as well as stochastic and nonstochastic generative models.
Single-network neural nets are composed of just two layers, like a cat's cradle. Multinetwork nets link double-layer nets together in long chains, each one functioning as a subnetwork in its own right.
Restricted Boltzmann machines are composed of binomial neurons; that is, they can only activated with the values one and zero. Continuous restricted Boltzmann machines accept decimals. That is the only difference.
An autoencoder is a neural network used for dimensionality reduction; that is, for feature selection and extraction. Autoencoders with more hidden layers than inputs run the risk of learning the identity function -- where the output simple equals the input -- thereby becoming useless.
Denoising autoencoders are an extension of the basic autoencoder, and represent a stochastic version of the autoencoder. Denoising autoencoders attempt to address identity-function risk by randomly corrupting input (i.e. introducing noise) that the autoencoder must then reconstruct, or denoise.