Simple neural network in OCaml

Train and run a neural network in OCaml.

Technical details

This implements a feed-forward, multi-layered and fully-connected network. Network learning is implemented as gradient descent using backpropagation algorithm.

How to use

Two methods are exposed (neuralnet.mli):

Neuralnet.fit - returns a trained neural network and expects the following arguments:

• dataset - list of tuples (input, output) where input/output are float arrays
• alg_type - optional. Gradient descent type: Batch, Stohastic or MiniBatch. Defaults to Batch.
• layers - optional. List of sizes of hidden layers in the network. Defaults to [] (no hidden layers).
• activation - optional. Activation function to be used. Defaults to Activation.sigmoid.
• max_iter - optional. Run the gradient descent up to max_iter iterations. Defaults to 10000.
• eps - optional. Run the gradient descent until learning error reaches eps. Defaults to 1e-3.
• rate - optional. Coefficient to multiply gradients with when updating weights. Defaults to 0.2.
• verbose - optional. Learning errors through iterations will be printed to stderr if set to true. Defaults to false.

Neuralnet.predict - runs the input through the network and returns the output. Expects following arguments:

• net - trained neural network. The thing Neuralnet.fit returns.
• x - an input to the network.

Example

See example.ml.

To run it do:

   make
   ./example