The sampling complexity of learning invertible residual neural networks