隐藏层的训练不会提高准确性

Question

我一直在对这个由 4 层组成的神经网络进行编程：

第一个有 2 个神经元，第二个有 2 个，第三个有 2 个，输出一个有一个神经元，我制作了这个模式来显示我想要重现的内容：

这是代码，您可以尝试运行它（python 3.7）：

import numpy as np

import matplotlib.pyplot as plt

#Calculus of the sigmoid

def sigmoid(z):

    return 1.0/(1+ np.exp(-z))

#Calculus of the sigmoid derivation

def sigmoid_derivative(y):

    return y * (1.0 - y)

#Initialisation of the class (input, output, targets, weights, biais)

class NeuralNetwork:

    def __init__(self, x, y):

        self.input      = x

        self.weights1   = np.random.rand(self.input.shape[1],2)

        self.weights2   = np.random.rand(2,2)

        self.weights3   = np.random.rand(2,2)

        self.weights4   = np.random.rand(2,1)

        self.y          = y

        self.output     = np.zeros(self.y.shape)

        self.bias1      = np.random.rand(1,2)

        self.bias2      = np.random.rand(1,2)

        self.bias3      = np.random.rand(1,2)

        self.bias4      = np.random.rand(1,1)

        self.learning_rate = 0.005

    #simple feed forward

    def feedforward(self):

        self.layer1 = sigmoid(np.dot(self.input, self.weights1) + self.bias1)

        self.layer2 = sigmoid(np.dot(self.layer1, self.weights2) + self.bias2)

        self.layer3 = sigmoid(np.dot(self.layer1, self.weights3) + self.bias3)

        self.output = sigmoid(np.dot(self.layer2, self.weights4) + self.bias4)

    #Back propagation algorithme

    def backprop(self):

        # application of the chain rule to find derivative of the loss function with respect to weights4, weights3, weights2, weights1 and the associated bias

        delta_4 = 2*(self.y - self.output) * sigmoid_derivative(self.output)

        d_weights4 = np.dot(self.layer3.T, delta_4)

        d_bias4 = delta_4

        d_bias4 = d_bias4.mean(axis=0)

        delta_3 = np.dot(delta_4, self.weights4.T) * sigmoid_derivative(self.layer3)

        d_weights3 = np.dot(self.layer2.T, delta_3)

        d_bias3    = delta_3

        d_bias3    = d_bias3.mean(axis=0)