作者提出了MILN: multilayer in-place learning network这个网络，
layer 4 和 layer2/3 是feature layer，其中layer 4 使用的是非监督学习，layer2/3 使用监督学习。（Laminar cortex: 层状皮层）The weight vector of each neuron is not computed based on gradient. Instead, it is the amnesic average (called
the lobe component) with properly scheduled, experience dependent plasticity.
如果输入y属于区域R，那么
就可以使均方误差
最小，其中
The main ideas are as follows. Each MILN is a generalpurpose regression network.
It develops by taking a series of input-output pairs, whenever the desired output vector is available. Otherwise, given any input without desired output,
it estimates the corresponding output vector
based on what it has learned so far. This is basically how MILN interleaves the training phases and testing phases in any way that is needed for effective development.
由时间t1到t2的转换时间 取决于网络发育的时长 以及 网络达到正常性能所需要的时间。 对于具体的工程系统(项目)，t1和t2可以根据网络的层数和每层神经元的个数来估计。