Lately, deep learning has seen extraordinary success in learning complex tasks involving natural signals such as images, speech, etc. Motivated by this, we begin the thesis with the exploration of a deep learning-based end to end communication system design. First, we propose a practical method for training auto encoder based communication systems in real channels without the knowledge of channel models. Next, after identifying that autoencoder models do not account for the noise in the system, we develop a deep learning-based framework based on Variational Inference which acknowledges the existence of corruption in channel. Further, the developed framework helps to quantify the noise as well as allow designers to incorporate domain knowledge into the training procedure. Exploring online learning, next we proceed to study blind beamforming for mmWave systems, which can significantly reduce the overhead in Multiple Input Multiple Output (MIMO) systems. For this problem, we develop a deep reinforcement learning-based solution for blind beam alignment in multi-user system with almost no overhead. Noting that efficient spectrum utilization is essential to meet the ambitious demands of future wireless and Cognitive Radio Networks (CRN) can be a good solution, we propose a two stage reinforcement learning based technique for the case of opportunistic spectrum access in cognitive radio networks. In the first stage, we rank the channel according to its chances of being free. In the second state, we use a Bayesian method to estimate residual off time and allow the SU to transmit until without further sensing for the estimated time, except when a collision happens. Finally, we explore algorithms that can adapt to changes in the environment as communication systems need to operate in constantly changing environments. In such scenarios, one-time learning is not adequate to improve the performance as the distributional assumptions underlying the learning methods could change. However, continuous retraining of such systems is also not an efficient option as it requires new data acquisitions and downtimes. In this context, we investigate learning techniques that can adapt to new environments promptly.