Exact priors of finite neural networks