Neural Network in Haskell

A simple neural network library in Haskell. The project aims to be of use to newcomers to Haskell, and an exploratory tool for neural networks. The entirety of this learning library is written from scratch, and hence may not be the best choice for efficiency.

Prerequisites

• Installation of GHCI is required. Instructions to install using GHCup (recommended) are available here (version - 9.4.7)

Usage

The Main.hs file provides two interfaces to the Neural Network's operations: train_model and predict_values.

Execute Main.hs in the terminal.

To train a model,

• Set hyperparameters in Params.hs. Note: Currently, a single activation is applied throughout the network.
• Execute :reload to update the compiled source (alternatively, :load Main.hs if compiling for the first time)
• Set the training features & their corresponding outputs in variables (here, xs and ys)
• Provide randomised weights corresponding to the network structure. For each layer of n inputs and m outputs, the corresponding weights must be a list of m lists, each of length n + 1. Note: Currently, the feature inputs are implicitly considered to be an input layer with no activation.
• Use function train_model to obtain the new weights of your network, providing the number of epochs required.

Predicting values is straightforward; predict_values uses the network's learnt weights to predict outputs for given data points.

Sample execution:

Multiple of 5

> ghci Main.hs
ghci> w = [ [[0.3, 0.4], [0.9, 0.6], [5, 0.3], [2, 1], [0.001, 0.5]], [[0.6, 0.5, 0.6, 0.3, 0.8, 0.9]]] -- a 1x5x1 network
ghci> xs = [[1], [2], [3], [4], [9], [8], [5], [3], [1], [0], [9], [8]]                              
ghci> ys = map (map (\x -> 5*x)) xs                                                                  
ghci> w2 = train_model w xs ys 50                                                                    
ghci> predict_values w2 xs                                                                           
[[4.9379373],[9.953882],[14.969826],[19.985771],[45.065495],[40.049545],[25.001715],[14.969826],[4.9379373],[0.17317355],[45.065495],[40.049545]]

for Params.hs:

activation :: Activation
activation = relu

activation' :: Activation_Prime
activation' = relu_prime

error_prime :: Y -> Y -> [Float]
error_prime = mse_prime

learning_rate :: Float
learning_rate = 0.001

XNOR

> ghci Main.hs
ghci> xs = [[0, 0], [0, 1], [1, 0], [1, 1]]
ghci> ys = [[1], [0], [0], [1]]
ghci>  w = [[[0.04, 0.5, 0.08], [0.9, 0.08, 0.7], [0.07, 0.9, 0.6]], [[0.7, 0.05, 0.06, 0.08]]] -- a 2-3-1 network
ghci> w2 = train_model w xs ys 500                      
ghci> predict_values w2 xs                              
[[0.9851537],[1.060858e-3],[6.017508e-3],[0.9790084]]

for Params.hs:

activation :: Activation
activation = tanh

activation' :: Activation_Prime
activation' = tanh_prime  

error_prime :: Y -> Y -> [Float]
error_prime = mse_prime  

learning_rate :: Float
learning_rate = 0.1