Modeling_Cellular_Perturbations_with_the_Sparse_Additive_Mechanism_Shift_Variational_Autoencoder_postNeurips

Figure 5: CPA-VAE represented as an generative process (left) and as a graphical model (right). A.1 Mini-batch optimization In this section, we provide a detailed description of how the ELBO is computed from mini-batches for optimization. During training, we iterate through shuffled versions of the training dataset and receive batches of indices B = {i1,...,i |B|}. Di,t be the total number of samples in the batch that have received perturbation t. Let P be a hyperparameter number of particles.