学习笔记：深度学习网络的优化-初始化参数

使用梯度下降法训练神经网络时，需要对参数初始化。对w参数(weight matrices) 和b参数(bias vectors)，使用不同的初始化策略会对训练结果和效率有非常大的影响：

1. Speed up the convergence of gradient descent 加速梯度下降的收敛速度
2. Increase the odds of gradient descent converging to a lower training (and generalization) error 使误差降低的概率更大

不同参数的方案：

1. 初始化为0
   • fails to "break symmetry" 未能破坏对成性
   • This means that every neuron in each layer will learn the same thing

for l in range(1, L):
    parameters['W' + str(l)] = np.zeros((layers_dims[l],layers_dims[l-1]))
    parameters['b' + str(l)] = np.zeros((layers_dims[l],1))
复制代码

2. 随机初始化
   • success to "break symmetry",make sure different hidden units can learn different things
   • Initializing weights to very large random values does not work well.
   • poor initialization can lead to vanishing/exploding gradients 导致梯度消失/梯度爆炸

for l in range(1, L):
        parameters['W' + str(l)] = np.random.randn(layers_dims[l],layers_dims[l-1]) * 10
        parameters['b' + str(l)] = np.zeros((layers_dims[l],1))
复制代码

3. 定制初始化方案——He initialization （author of He et al., 2015）

for l in range(1, L + 1):
        parameters['W' + str(l)] = np.random.randn(layers_dims[l],layers_dims[l-1]) * np.sqrt(2/layers_dims[l-1])
        parameters['b' + str(l)] = np.zeros((layers_dims[l],1))
复制代码

总结：

• Different initializations lead to different results
• Random initialization is used to break symmetry and make sure different hidden units can learn different things
• Don't intialize to values that are too large
• He initialization works well for networks with ReLU activations.