AT Results and Proofs

The result follows. 1 B Gradient Derivations B.1 Weights gradient First, we compute the gradient with respect to weights vector w R B.2 Location gradients Here we take the gradient with respect to a single pseudopoint u Increasing this number is typically expensive to obtain in practice. Bayesian Logistic Regression experiment presented in Section 4. Posterior approximation metrics, coreset gradients and learning rates Estimation of differential privacy cost at all experiments was based on TensorFlow privacy implementation of moments accountant for the subsampled Gaussian mechanism. Experiments were performed on a CPU cluster node with a 2x Intel Xeon Gold 6142 and 12GB RAM.