The quantum neural network is one of the promising applications for near-term noisy intermediate-scale quantum computers.  A quantum neural network distills the information from the input wavefunction into the output qubits.  In this talk, we argue that this process can also be viewed from the opposite direction: the quantum information in the output qubits is scrambled into the input.  Based on this idea, in the first work, we find strong correlation between the dynamical behavior of the tripartite information and the loss function in the training process, from which we identify that the training process has two stages for randomly initialized networks. In the second work, we further use this idea to construct quantum neural network structures that is the most efficient for general training process.