Following up from Seminar 1, we revisit the design of end to end communication system using deep learning. Most of the neural network architectures in current literature leverage autoencoders to design the encoder at the transmitter and decoder at the receiver and train them jointly by modeling transmit symbols as latent codes from the encoder. However, in communication systems, the receiver has to work with noise corrupted versions of transmit symbols. Traditional autoencoders are not designed to work with latent codes corrupted with noise and training such a model without restrictions will cause the transmit energy to increase unboundedly. Previous works overcame this difficulty by using an additional normalization layer at the output of the transmitter, which hard constraints the maximum transmit power. Noting this gap in existing literature, we develop a framework to design end to end communication systems that account for the existence of noise corrupted transmit symbols. Instead of introducing a hard constraint in the model, we propose a framework for deriving alternate objective functions based on the concepts of variational inference. The new terms introduced by the proposed framework soft limits the transmit power implicitly. Further, domain knowledge such as channel type can be systematically integrated into the objective through the framework. Further, we show how to derive objective functions for training models in multiple popular channel models. Simulation results comparing the proposed approach with the existing works along with failure modes and interesting observations will be discussed.