通过神经网络回归最大函数

Question

我正在训练自己学习神经网络。有一个功能，我无法让我的神经网络学习：f(x) = max(x_1, x_2)。它似乎是一个非常简单的功能，有2个输入和1个输入但是3层神经网络训练了超过一千个样本和2000个时期使它完全错误。我正在使用deeplearning4j。

有什么理由为什么max函数对于神经网络来说很难学习，或者我只是调错了？

Answer 1

如果你将x1和x2限制在一个区间内，例如，这不是很难。在[0,3]之间。从deeplearning4j示例中取出“RegressionSum”示例，我很快将其重写为学习max而不是总和，并且它可以很好地给我结果：

Max(0.6815540048808918,0.3112081053899819) = 0.64
Max(2.0073597506364407,1.93796211086664) = 2.09
Max(1.1792029272560556,2.5514324329058233) = 2.58
Max(2.489185375059013,0.0818746888836388) = 2.46
Max(2.658169689797984,1.419135581889197) = 2.66
Max(2.855509810112818,2.9661811672685086) = 2.98
Max(2.774757710538552,1.3988513143140069) = 2.79
Max(1.5852295273047565,1.1228662895771744) = 1.56
Max(0.8403435207065576,2.5595015474951195) = 2.60
Max(0.06913178775631723,2.61883825802004) = 2.54

下面是我修改后的版本的RegressionSum示例，该版本最初来自Anwar 3/15/16：

public class RegressionMax {
    //Random number generator seed, for reproducability
    public static final int seed = 12345;
    //Number of iterations per minibatch
    public static final int iterations = 1;
    //Number of epochs (full passes of the data)
    public static final int nEpochs = 200;
    //Number of data points
    public static final int nSamples = 10000;
    //Batch size: i.e., each epoch has nSamples/batchSize parameter updates
    public static final int batchSize = 100;
    //Network learning rate
    public static final double learningRate = 0.01;
    // The range of the sample data, data in range (0-1 is sensitive for NN, you can try other ranges and see how it effects the results
    // also try changing the range along with changing the activation function
    public static int MIN_RANGE = 0;
    public static int MAX_RANGE = 3;

    public static final Random rng = new Random(seed);

    public static void main(String[] args){

        //Generate the training data
        DataSetIterator iterator = getTrainingData(batchSize,rng);

        //Create the network
        int numInput = 2;
        int numOutputs = 1;
        int nHidden = 10;
        MultiLayerNetwork net = new MultiLayerNetwork(new NeuralNetConfiguration.Builder()
                .seed(seed)
                .iterations(iterations)
                .optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT)
                .learningRate(learningRate)
                .weightInit(WeightInit.XAVIER)
                .updater(Updater.NESTEROVS).momentum(0.9)
                .list()
                .layer(0, new DenseLayer.Builder().nIn(numInput).nOut(nHidden)
                        .activation("tanh")
                        .build())
                .layer(1, new OutputLayer.Builder(LossFunctions.LossFunction.MSE)
                        .activation("identity")
                        .nIn(nHidden).nOut(numOutputs).build())
                .pretrain(false).backprop(true).build()
        );
        net.init();
        net.setListeners(new ScoreIterationListener(1));


        //Train the network on the full data set, and evaluate in periodically
        for( int i=0; i<nEpochs; i++ ){
            iterator.reset();
            net.fit(iterator);
        }

        // Test the max of some numbers (Try different numbers here)
        Random rand = new Random();
        for (int i= 0; i< 10; i++) {
            double d1 = MIN_RANGE + (MAX_RANGE - MIN_RANGE) * rand.nextDouble();
            double d2 =  MIN_RANGE + (MAX_RANGE - MIN_RANGE) * rand.nextDouble();
            INDArray input = Nd4j.create(new double[] { d1, d2 }, new int[] { 1, 2 });
            INDArray out = net.output(input, false);
            System.out.println("Max(" + d1 + "," + d2 + ") = " + out);
        }

    }

    private static DataSetIterator getTrainingData(int batchSize, Random rand){
        double [] max = new double[nSamples];
        double [] input1 = new double[nSamples];
        double [] input2 = new double[nSamples];
        for (int i= 0; i< nSamples; i++) {
            input1[i] = MIN_RANGE + (MAX_RANGE - MIN_RANGE) * rand.nextDouble();
            input2[i] =  MIN_RANGE + (MAX_RANGE - MIN_RANGE) * rand.nextDouble();
            max[i] = Math.max(input1[i], input2[i]);
        }
        INDArray inputNDArray1 = Nd4j.create(input1, new int[]{nSamples,1});
        INDArray inputNDArray2 = Nd4j.create(input2, new int[]{nSamples,1});
        INDArray inputNDArray = Nd4j.hstack(inputNDArray1,inputNDArray2);
        INDArray outPut = Nd4j.create(max, new int[]{nSamples, 1});
        DataSet dataSet = new DataSet(inputNDArray, outPut);
        List<DataSet> listDs = dataSet.asList();
        Collections.shuffle(listDs,rng);
        return new ListDataSetIterator(listDs,batchSize);

    }
}

Answer 2

只是想指出：如果你使用relu而不是tanh而不是实际上有一个确切的解决方案，我想如果你将网络缩小到这个完全相同的大小（1个隐藏层有3个节点），你总是会结束使用这些权重（节点的模块排列和权重的缩放（第一层由伽马缩放，第二层为1 / gamma））：

max(a,b) = ((1, 1, -1)) * relu( ((1,-1), (0,1), (0,-1)) * ((a,b)) )

其中*是矩阵乘法。

此等式将以下人类可读版本转换为NN语言：

max(a,b) = relu(a-b) + b = relu(a-b) + relu(b) - relu(-b)

我实际上没有对它进行测试，我的观点是，从理论上说，网络学习这个功能应该很容易。

编辑：我刚测试了这个，结果就像我预期的那样：

[[-1.0714666e+00 -7.9943770e-01  9.0549403e-01]
 [ 1.0714666e+00 -7.7552663e-08  2.6146751e-08]]

和

[[ 0.93330014]
 [-1.250879  ]
 [ 1.1043695 ]]

其中对应的第一层和第二层。转置第二个并乘以第一组权重，最终得到一个标准化版本，可以很容易地与我的理论结果进行比较：

[[-9.9999988e-01  9.9999988e-01  1.0000000e+00]
 [ 9.9999988e-01  9.7009000e-08  2.8875675e-08]]