我的单层感知器不工作

Question

这是代码。

public class Adaline
{
    private int _layer;
    public int Layer { get { return _layer; } }
    private int _epoch;
    public int Epoch { get { return _epoch; } }
    private double _error;
    public double Error { get { return _error; } }

    private double[] _weights;

    public Adaline(int layer)
    {
        _layer = layer;
        _weights = new double[layer];
        Reset();
    }

    public void Reset()
    {
        Random r = new Random();
        for (int i = 0; i < _layer; i++)
            _weights[i] = r.NextDouble() - 0.5;
        _error = 1;
    }

    public void Train(BasicTrainSet<double> trainset, double learnRate)
    {
        double ers = 0;
        for(int p = 0; p < trainset.DataCount; p++)
        {
            double result = Compute(trainset.Input[p], true);
            double error = trainset.Output[p] - result;

            for (int i = 0; i < _weights.Length; i++)
            {
                _weights[i] += error * trainset.Input[p][i] * learnRate;
            }

            ers += Math.Abs(error);
        }
        _epoch++;
        _error = ers;
    }

    public double Compute(double[] input, bool quan)
    {
        double result = 0;
        for (int i = 0; i < _layer; i++)
            result += Math.Tanh(_weights[i] * input[i]);
        //double result = _weights.Zip(input, (a, b) => Math.Tanh(a * b)).Sum();
        return quan ? (result >= 0 ? 1 : 0) : result;
    }
}

当我尝试像这样训练和门时，它是这样工作的。 Up four results are from this code 这很奇怪，因为算法没有任何问题。权重越来越大。我哪里弄错了？

Answer 1

在计算每个神经元输出的代码中，您没有正确应用激活函数。您需要找到权重和每个神经元的输入之间的点积然后然后应用激活函数。您在每次加权累加后应用激活函数，这是不正确的。

积累，然后应用激活函数：

public double Compute(double[] input, bool quan)
{
    double result = 0;
    for (int i = 0; i < _layer; i++)
        result += _weights[i] * input[i]; // Change - accumulate first
    result = Math.Tanh(result); // Change - now apply activation function

    return quan ? (result >= 0 ? 1 : 0) : result;
}

我的单层感知器不工作

My single layer perceptrone is not working

c#

machine-learning

perceptron

neural-network