Что такое обратная функция активации(inverse activation function)?

It’s more chaotic than a traditional neural network due to the sparse connectivity and skip connections, but it is more explainable since the activation functions and structure have been explicitly optimised.

The most striking feature of the network is the x parameter (position of the cart) has an almost direct line to the output node, which calculates force, intercepted only by a single hidden neuron with an inverse activation function. This line has the effect of providing a force towards the center of the screen regardless of the sign of the shared weight value: the network has learned how to get the cart to the center in a weight-agnostic way.