TCT: Convexifying Federated Learning using Bootstrapped Neural Tangent Kernels

State-of-the-art federated learning methods can perform far worse than their centralized counterparts when clients have dissimilar data distributions.