sams-neurons/sam-neurons-words.js

class NeuralNetwork {
    constructor(inputNodes, hiddenNodes, outputNodes) {
        this.inputNodes = inputNodes;
        this.hiddenNodes = hiddenNodes;
        this.outputNodes = outputNodes;

        // Initialize weights with random values between -1 and 1
        this.weights_ih = this.initializeWeights(this.hiddenNodes, this.inputNodes);
        this.weights_ho = this.initializeWeights(this.outputNodes, this.hiddenNodes);

        // Biases
        this.bias_h = this.initializeBiases(this.hiddenNodes);
        this.bias_o = this.initializeBiases(this.outputNodes);

        // Learning rate
        this.learningRate = 0.1;
    }

    // Helper method to initialize weights
    initializeWeights(rows, cols) {
        return Array(rows).fill().map(() => Array(cols).fill().map(() => Math.random() * 2 - 1));
    }

    // Helper method to initialize biases
    initializeBiases(size) {
        return Array(size).fill().map(() => Math.random() * 2 - 1);
    }

    // Sigmoid activation function
    sigmoid(x) {
        return 1 / (1 + Math.exp(-x));
    }

    // Sigmoid derivative (used for backpropagation)
    sigmoidDerivative(x) {
        return x * (1 - x);
    }

    // Feedforward function
    feedforward(inputArray) {
        // Calculate hidden layer outputs
        let hidden = this.weights_ih.map((row, i) => 
            this.sigmoid(row.reduce((sum, weight, j) => sum + weight * inputArray[j], 0) + this.bias_h[i])
        );

        // Calculate output layer outputs
        let output = this.weights_ho.map((row, i) => 
            this.sigmoid(row.reduce((sum, weight, j) => sum + weight * hidden[j], 0) + this.bias_o[i])
        );

        return output;
    }

    // Train the network
    train(inputArray, targetArray) {
        // Feedforward
        let hidden = this.weights_ih.map((row, i) => 
            this.sigmoid(row.reduce((sum, weight, j) => sum + weight * inputArray[j], 0) + this.bias_h[i])
        );
        let outputs = this.weights_ho.map((row, i) => 
            this.sigmoid(row.reduce((sum, weight, j) => sum + weight * hidden[j], 0) + this.bias_o[i])
        );

        // Calculate the output errors
        let outputErrors = targetArray.map((target, i) => target - outputs[i]);

        // Calculate output gradients
        let outputGradients = outputs.map((output, i) => 
            this.sigmoidDerivative(output) * outputErrors[i] * this.learningRate
        );

        // Calculate hidden errors
        let hiddenErrors = this.weights_ho[0].map((_, i) => 
            this.weights_ho.reduce((sum, row) => sum + row[i] * outputErrors[0], 0)
        );

        // Calculate hidden gradients
        let hiddenGradients = hidden.map((h, i) => 
            this.sigmoidDerivative(h) * hiddenErrors[i] * this.learningRate
        );

        // Update weights and biases
        this.weights_ho = this.weights_ho.map((row, i) => 
            row.map((weight, j) => weight + outputGradients[i] * hidden[j])
        );
        this.weights_ih = this.weights_ih.map((row, i) => 
            row.map((weight, j) => weight + hiddenGradients[i] * inputArray[j])
        );
        this.bias_o = this.bias_o.map((bias, i) => bias + outputGradients[i]);
        this.bias_h = this.bias_h.map((bias, i) => bias + hiddenGradients[i]);
    }
}

// Example usage:

// Example dataset (XOR problem)
let strings = ["hello", "neural", "world", "how are", "you", "?"]

let inputs = [
    [0, 0, 0, 0, 0 ,0],
    [1, 0, 0, 0, 0, 0],
    [1, 0, 1, 0, 0, 0],
    [1, 1, 1, 0, 0, 0],
    [1, 1, 1, 0, 1, 0],
    [1, 1, 1, 0, 1, 1],
    [1, 1, 1, 1, 1, 1]
];

let nn = new NeuralNetwork(strings.length, inputs.length, 1);  // Increased hidden nodes to 4

let targets = [
    [0],
    [1],
    [2],
    [3],
    [4],
    [5],
    [6]
];



// Example dataset (XOR problem)
let strings2 = ["i'm", "fine", "myself."]

let inputs2 = [
    [0, 0, 0],
    [1, 0, 0],
    [1, 0, 1],
    [1, 1, 1]
];

let nn2 = new NeuralNetwork(strings2.length, inputs2.length, 1);  // Increased hidden nodes to 4

let targets2 = [
    [0],
    [1],
    [2],
    [3]
];


// Training the neural network
function train(nn, inputs, targets) {
    if (targets.length != inputs.length) {
        throw new Error("You dummy, targets length should be equal to inputs length.")
    }
    for (let i = 0; i < 50000; i++) {  // Increased training iterations
        let index = Math.floor(Math.random() * 4);
        nn.train(inputs[index], targets[index]);
    }
}


function think(nn, feeds, strings) {
    if (strings.length != feeds[0].length) {
        throw new Error("You dummy, strings array length should be equal to a feed input length.")
    }
    feeds.forEach((item, index) => {
        let feededVal = Math.round(nn.feedforward(item))
        console.log([feededVal, item])
        if (Math.round(feededVal)  == 1) {
            let sum = 0;
            item.forEach(num => sum += num);
            try { console.log(strings[sum - 1]) } catch {}
        }
    })
}

train(nn, inputs, targets)
think(nn, inputs, strings)

train(nn2, inputs2, targets2)
think(nn2, inputs2, strings2)