Machine Learning Unsupervised Backpropagation

Question

I'm having trouble with some of the concepts in machine learning through neural networks. One of them is backpropagation. In the weight updating equation,

delta_w = a*(t - y)*g'(h)*x

t is the "target output", which would be your class label, or something, in the case of supervised learning. But what would the "target output" be for unsupervised learning?

Can someone kindly provide an example of how you'd use BP in unsupervised learning, specifically for clustering of classification?

Answer 1

Backpropagation in unsupervised learning; probably the models like RBM's(restricted boltzmann machines) and Autoencoders can be used for the problem you mentioned above.
RBMs are stochastic networks that are commonly represented as a graphical model with connectivity limitations. There is no output layer in this configuration, and the relationship between the input and latent layers is bidirectional, similar to an undirected graphical model. It is attempting to learn a distribution of inputs (observed and unobserved variables)
Autoencoders are simply feedforward neural networks with weights tuned in such a way that the output is a reconstructed input (optionally sparse activations can be stored in optimization

function). There can be more hidden layers, but for a better starting point, the weight initialization employs greedy layer wise training. As a result, the target function will be input itself in order to respond to the inquiry.
In order to train an artificial neural network, backpropagation is utilised to compute the derivatives of the error function with respect to the network's weights. The "errors" "propagate" backwards through the network, hence the name.