Neural Network

A flexible all purpose multilayer perceptron neural network class.

The NeuralNetwork class handles generating network parameters, feedforward evaluation, loss metrics, single or batch training with gradient descent, and stores input loss gradients for easy chaining.

/* Create Neural Network Instance */

NeuralNetwork myNeuralNetwork(
    3, // input node count
    {
        HiddenLayerParameters(10, RELU), // hidden layer with 10 nodes and relu activation
        HiddenLayerParameters(16, TANH), // hidden layer with 16 nodes and tanh activation
        
        HiddenLayerParameters(4, LINEAR) // output layer with 4 nodes and linear activation
    },
    SOFTMAX, // normalization function
    CATEGORICAL_CROSS_ENTROPY // loss function
);

/* Initialize Parameters */

// optionally set distribution range for initial weights and bias
myNeuralNetwork.initializeRandomLayerParameters();

/* Use DataPoint Class For Training */

std::vector<DataPoint> trainingDataPoints;

trainingDataPoints.emplace_back(
    Matrix({{ 0.2 }, { -0.4 }, { 0.3 }}), // input
    Matrix({{ 0.01 }, { 0.97 }, { 0.01 }, { 0.01 }}) // expected output
);

// ... add other training data ...

/* Train Using A Single DataPoint */

float learningRate = 0.05;

myNeuralNetwork.train(trainingDataPoints[0], learningRate);

/* Train Using A Batch */

float learningRate = 0.1;

myNeuralNetwork.batchTrain(trainingDataPoints, learningRate);

/* Make A Prediction */

Matrix input({{ 0.3 }, { 0.7 }, { -0.2 }});

Matrix output = myNeuralNetwork.calculateFeedForwardOutput(input);

/* Calculate The Loss For A Known Point */

Matrix expectedOutput({{ 0.97 }, { 0.01 }, { 0.01 }, { 0.01 }});

float loss = myNeuralNetwork.calculateLoss(input, expectedOutput);