Mausr
Marek's Unicode Symbol Recognizer
Mausr is a neural network library written from scratch and used for recognition of hand written unicode symbols. The library is relatively general and written with emphasis on datastructures, and extensibility (and maybe a little of performance, too :).
The main reason behind this project was my personal interest in neural networks. I decided to create unicode symbol recognizer because I often find myslef googling for an unicode symbol and it takes too much time.
Author: Marek Fiser < mausr@marekfiser.cz >
Running instance: http://www.mausr.com/
License: The MIT license, see LICENSE.txt for details.
Main features of neural net library
- Basic neural network layout with input layer, output layer, and any number of hidden layers.
- Extensible neuron activation function, stadnard sigmoid function implemented.
- Extensible net cost function, standard logistic regression cost function implemented.
- Extensible gradient based optimization algorithms with visual and algorithmical tests, implemented four:
- Basic gradient descent,
- Gradient descent with momentum, and
- RProp+ algorithm.
- iRProp- algorithm.
- Back-propagation learning algorithm.
- Efficient, vectorized, and paralellized implementation.
- Regularization implemented to avoid overfitting.
- Contains around 25 unit tests that ensure correctness of core components of training and evaluation algorithms.
- Also contains simple visual tests of optimization algorithms to ensure expected behavior.
Main features of web interface
- Search for hand drawn symbols using canvas.
- Interface for neural network settings and training.
- Includes real-time visual feedback of traind and test validation erors using signal-r and google chart API.
- Interface for training of new symbol drawings.
- Interface for approving anonymously submitted trainng data.
- Database for storing symbols, drawings, and users.