what is the mean of "Classifiying Patterns" in Neural network?

maybe, this is general question, i dont know. But i am working neural network and i dont understand somethings on "Essence of Neural Networks" book.

"Classifiying Pattern" 
"Clustering Patterns"
"Pattern Association"

In this book, is the pattern means the training开发者_StackOverflow社区 input? First time i am working about Neural Network, this book is available for me?

The term pattern is used in the context of neural networks to mean a set of activations across a pool of units (neurons).

These are all different tasks involving patterns:

"Classifiying Pattern"

Getting a net to answer questions of the form: is input x of type a,b, or c?

An example problem would something like: given a picture (input pattern) of an animal, output its species (output pattern)

"Clustering Patterns"

Getting a net to answer questions of the form: whats an reasonable way of subdividing input x or whats a good way to group inputs {x1,x2,x3,...}

"Pattern Association"

Getting a net to map some input to some output

Imagine trying to get a network to solve the standard XOR problem: you give the net two bits and want it to output the XOR of them (e.g. {0,1}->{1}, {1,1}->{0}, etc). In neural network terms, you need it to associate each input pattern ({0,1}) with the correct output pattern ({1}).

In this book, is the pattern means the training input?

A pattern is related to the input, as in: there is a pattern in your input and you're trying to find it. There is a pattern in your training input, which you want to learn, but you're trying to create a "general" solution which not only recognizes a pattern in your training input but it also recognizes a pattern in your validation and testing input. The assumption is that in a "perfect world," if you're predicting the validation and testing patters correctly, then you should be able to predict a pattern any input that comes your way.

In LeCun "A theoretical framework for backpropagation" 1988, ten years older than the book you mentioned, the pattern is an instance of the input data that defines the value of each layer of the network.

It is used interchangeably for the indexing of a training dataset. For each pattern $p=1,...P$, the training dataset is indicated with ${I_p, D_p}$, where I_p is the input of the NN and $D_p$ is the desired output.

In my view, it makes sense to read pattern as input data configuration, or input data event in an optimal control perspective.