Designing Universally-Approximating Deep Neural Networks: A First-Order Optimization Approach | Synapse