As autonomous driving technologies have been dominated rapidly, machine learning (ML) has become a technology they could never survive without when designing a safe and efficient autonomous driving vehicle. In this paper, we provide a comprehensive review of machine learning use cases in autonomous driving, cover the current advancement, discuss the key challenges and look forward to the future. The three principal ML-enabled parts of an autonomous vehicle perception, planning and control modules are analysed, and their development is explained in terms of how deep learning architectures, in particular convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have enabled capabilities of enhanced performance. However, our research shows that object detection and classification systems perform as high as 95% accuracy but still have their own challenges regarding adverse weather conditions and complex traffic scenarios. Finally, we outline critical technical challenges in realizing real-time processing, high data quality requirements, and reliable models and describe our novel solutions through hybrid architectures and optimized training strategies. I show performance evaluations across multiple autonomous driving platforms and that ML-based systems beat traditional rule-based approaches at 40% in terms of decision-making accuracy and 35% in terms of response time. Moreover, the paper deliberates on vital ethical and regulatory concerns, arguing for uniform safety procedures and clear ways of making decisions. We find our results show how emerging technologies like federated learning and edge computing can help move the needle on autonomous driving. This research presents a structured framework of ML's role in autonomous driving, the future research direction, and potential industry application with ML is explored.