Machine Learning for Low-Latency Communications presents the principles and practice of various deep learning methodologies for mitigating three critical latency components: access latency, transmission latency, and processing latency. In particular, the book develops learning to estimate methods via algorithm unrolling and multiarmed bandit for reducing access latency by enlarging the number of concurrent transmissions with the same pilot length. Task-oriented learning to compress methods based on information bottleneck are given to reduce the transmission latency via avoiding unnecessary data transmission. Lastly, three learning to optimize methods for processing latency reduction are given which leverage graph neural networks, multi-agent reinforcement learning, and domain knowledge. Low-latency communications attracts considerable attention from both academia and industry, given its potential to support various emerging applications such as industry automation, autonomous vehicles, augmented reality and telesurgery. Despite the great promise, achieving low-latency communications is critically challenging. Supporting massive connectivity incurs long access latency, while transmitting high-volume data leads to substantial transmission latency.