Restoration of convex CPL-functions by neural nets over RELU bases
Abstract:
The paper considers the issue of representing convex piecewise linear functions in the form of neural networks with activation functions RELU type and neural networks with a single non-linearity - so called the max-pooling layer. This question is in support of proof of a more general representation problem for any piecewise linear function in the form of a neural network over these bases. This problem is also solved, but so far the results obtained are under review, so it was decided to make emphasis on an intermediate, but no less important result. Also in work numerical characteristics are given that evaluate the quality of the obtained neural networks - non-linear complexity and depth.
