Convergence of Implicit Gradient Descent for Training Two-Layer Physics-Informed Neural Networks | Synapse