DigitRecognition

Very simple feed-forward neural network

It is trained and tested using the MNIST dataset of handwritten digits.

Building

Prerequisites: Gradle and JDK

gradle build

Build and run the jar with

gradle fatJar
java -jar build/libs/DigitRecognition-1.0.jar

Running

gradle run --args draw The possible arguments are draw, eval and train.

• draw creates a windows with a canvas for drawing numbers and having them classified
• train trains the network with the training data
• eval evaluates the performance of the network based on test data

The network is exported to and imported from ./network/{customWeights,customBiases}.txt

License

©2015 - 2019 Daniel Schäfer, BSD 3-Clause