Package org.shapelogic.machinelearning

Machine Learning

See:
Description

Class Summary

ConfidenceArraySelector	ConfidenceArraySelector takes a double[] and creates a String. Translates a double[] that could come from a neural network to either the number of the one that is winning walue if any is or to a name for that.
ExampleNeuralNetwork	ExampleNeuralNetwork for use in default neural network.
FFNeuralNetwork	Feed Forward Neural Network with external training.
FFNeuralNetworkStream	FFNeuralNetworkStream a feed forward neural network wrapped in a stream. Streams involved: ArrayOutputListStream: A list of names of input features that need to be streams FFNeuralNetwork: Creates stream of double[] from the neural network ConfidenceArrayListStream: Creat result stream of String Which of these needs names? None of them. It seems like the last stream will have the same output as the overall stream is the overall then really needed? No. There should probably be a stack of Streams. This could possibly be a LISP list, no there is no advantage to this. The whole think should be created lazily, but when it gets created, you can directly pipe one into the next. What should trigger the setup? As long as the first have a lazy setup we should be fine and it does. So the first should be any type of ListStream, while the next have to be ListCalcStream1 or any ListStream with one input.
FFNeuralNetworkWeights	FFNeuralNetworkWeights represent a trained feed forward neural network.
FFNeuralNetworkWeightsParser	FFNeuralNetworkWeightsParser parses FFNeuralNetworkWeights. I need a lookahead for the block line.

Package org.shapelogic.machinelearning Description

Machine Learning

This package currently contains a simple feed forward neural network integrated into the lazy stream framework.

The training of the neural network is done externally, but run internally.
Steps in supervised learning for the color particle analyzer.
To train the neural network you need a list of images each with one or more particles, here are some of the things you need to do:

• You have to assign different categories to the different particles
• You need to select which geometric property data streams you want as input
• This should be a minimal list otherwise you can get over fitting problems
• Possibly create new geometric property data streams
• Find out how many neural network layers you need
• Find out how many nodes in each layer you need
• Store this information so it can be reused

The reason for not doing training inside ShapeLogic is: ImageJ has the simple and powerful PlugInFilter concept and it would be very hard to get all that functionality into a PlugInFilter without creating something that is quite complex.

Joone - Java Object Oriented Neural Engine http://www.jooneworld.com/ is a great tool to train the neural network used in ShapeLogic.

Dependency

• This package can depend on util, calculation, streams.
• There is no dependency of any image processing code or any external libraries outside Java 6.
• This can be used in a Java application, just by adding shapelogic_.jar.