PP: Neural ordinary differential equations

Instead of specifying a discrete sequence of hidden layers, we parameterize the derivative of the hidden state using a neural network. 

Before: a discrete sequence of hidden layers.

After: the derivative of the hidden state.

Traditional methods: residual networks, RNN decoders, and normalizing flows build complicated transformations by composing a sequence of transformations to a hidden state.

we parameterize the continuous dynamics of hidden units using an ordinary differential equation (ODE) 常微分函数.

将h(t) 看作一个函数，可以用一个neural network学习h(t)的分布，然后输入层h(0) ----> 输出层h(T);