Neural-classifier

neural-classifier is a neural network library based on the first chapters from this book. It is divided on two systems: neural-classifier which is a general API for neural networks and neural-classifier/mnist which contains helper functions for working with MNIST/EMNIST datasets. For API documentation visit this page.

How to work with MNIST dataset?

• Unpack files in mnist/dataset directory.
• Load neural-classifier/mnist system: (ql:quickload :neural-classifier/mnist).
• Eval (neural-classifier-mnist:load-mnist-database) (this will take about 10-15 seconds).
• Create a neural network: (defparameter *nn* neural-classifier-mnist:make-mnist-classifier 35) where 35 is a number of hidden neurons.
• Execute (neural-classifier-mnist:train-epochs *nn* 10) to train the network for 10 epochs. This function will return data about the network's accuracy for each epoch.
• To test your own digits convert them to 784x1 matrix of type magicl:matrix/single-float and pass it to neural-classifier:calculate function.
• Also you can play with some other hyper-parameters, not only the number of hidden neurons. neural-classifier:*learn-rate* is how fast gradient descent algorithm works (i.e. how fast your network learns), neural-classifier:*decay-rate* is related to regularization and should be about 5/N where N is a number of training samples. Zero means no regularization.

How to build custom nets and data?

See GH pages for this project (link above). In general you need to write functions which translate your data and labels into magicl:matrix/single-float matrices. Then you create a net with neural-classifier:make-neural-network function and snakes generator which returns conses in the form (DATA . LABEL). To train a network for one epoch you call (neural-classifier:train-epoch).

Dependencies

• blas and lapack foreign libraries.
• magicl for matrix operations.
• nibbles for loading MNIST data.

magicl and nibbles can be downloaded with quicklisp.

What if the network shows good accuracy but fails to recognize my own digits?

If the accuracy returned by train-epochs is good, but the network fails to recognize digits draws by your own hand, try EMNIST database instead of MNIST. Copy four emnist-digits-* files to your MNIST directory preserving the name of destination files. Images in EMNIST set are transposed (x and y coordinates swapped), so do the same with your own images.