INPUT+ONE HIDDEN LAYER+OUTPUT 已经可以实现任意布尔函数。其实也可以是任意逼近 measurable函数。

MLP是universal approximators。既然One hidden layer的拟合能力已经这么强，可以让网络更宽一些。

Numerous and significant advantages accrue from using a flat net, including rapid quadratic optimization in the learning of weights, simplification in hardware as well as in computational procedures, and uniform system architecture for all four functionalities.

Functional-Link Net Computing: Theory, System Architecture, and Functionalities