Network Anomaly detection is rapidly growing field of information security due to its importance for protection of information networks. Being the first line of defense for network infrastructure, intrusion detection systems are expected to dynamically adapt with changing threat landscape. Deep learning is an evolving sub-discipline of machine learning which has delivered breakthroughs in different disciplines including natural language processing, computer vision and image processing to name a few. The successes of deep learning in aforementioned disciplines condone investigation of its application for solution of information security problems.This research aims at investigating deep learning approaches for anomaly-based intrusion detection system. In this study we propose, implement, evaluate and compare the use of Deep learning both as a refined representation learning mechanism as well as a new supervised classification mechanism for enhanced anomaly detection. Contributions of this research include Deep Supervised Learning and Deep Representation Learning for Network anomaly detection systems. For Deep Supervised Learning, anomaly detection models were developed by employing well-known deep neural network structures on both balanced and imbalanced datasets. For balanced Datasets we used four partitions of NSLKDD dataset while UNSWNB15 and ISCX2012 were employed as imbalanced datasets both of which contain 4.9% anomalous sample on average. For comparisons, conventional machine learning-based anomaly detection models were developed using well-known classification techniques. Both deep and conventional machine learning models were evaluated using standard model evaluation metrics. Results showed that DNN based anomaly detectors showed comparable or better results for detection of network anomalies. Deep Representation Learning involves using Deep learning to create efficient and effective Data representations from raw and high-dimensional network traffic data for developing anomaly detectors. Creating efficient representations from large volumes of network traffic to develop anomaly detection models is a time consuming and resource intensive task. Deep learning is proposed to automate feature extraction task in collaboration with learning subsystem to learn hierarchical representations which can be used to develop enhanced data driven anomaly detection systems. Four representation learning models were trained using well-known Deep Neural Network architectures to extract Deep representations from ISCX 2012 traffic flows. Each of these Deep representations is used to train anomaly detection models using twelve conventional Machine Learning algorithms to compare the performance of aforementioned deep representations with that of a human-engineered representation. The comparisons were performed using well known classification quality metrics. Results showed that Deep Representations perform comparable or better than human-engineered representations but require fraction of cost as compared to human-engineered representations due to inherent support of GPUs. Hyperparameter optimization of deep neural network used for current study is performed using Randomized Search. Experimental results of current research showed that Deep Neural Networks are an effective alternative for both representation learning and classification of network traffic for developing contemporary anomaly detection systems.