Wired or wireless, connectivity has been a vital commodity of life and more so recently in the realm of information age. Those who have access to faster, more reliable and more ubiquitous connectivity—put simply, those who are “better connected”—will have significant advantages in commerce, research and a host of other arenas. In regards to wireless communications, due to the explosion in demand for higher capacity networks, availability of free spectrum resources have become increasingly scarce. The UHF spectrum band in particular, due to its excellent electromagnetic properties, has been reported as inefficiently used and congested by many spectrum regulators of the world. This spectrum resource scarcity issue combined with the ongoing research and development for more intelligent, autonomous and self-aware radio communication led to a vast amount of research on the concept of Cognitive Radio. This thesis researches the learning unit of cognitive radios. The learning unit is responsible for processing information and autonomous decision making. In particular, the research is focused on the extraction and usage of contextual information from the radio environment (e.g. Radio Access Technology type, channel access pattern learning/recognition) and how such information could be exploited to improve the performance of the cognitive radio. The key metrics discussed will be based on information extraction under noise, channel blocking and interference reduction to primary users. We present a set of novel works involving Machine Learning, which is a branch of Artificial Intelligence. New implementation and use cases of state-of-the-art machine learning algorithms are presented that learn from real-life data. In a testbed setup we program software defined radios to recognize different Radio Access Technologies and their channel access patterns. The main technique used in the majority of the thesis is Artificial Neural Networks, concretely: Multi-Layer Perceptron Neural Nets, Self-Organizing Neural Nets, and Deep Auto-Encoders. In some of the works these neural network architectures have been combined in a novel way with Support Vector Machines, and Reinforcement Learning algorithms for channel classification and access. In this thesis we show that it is possible to achieve 95% correct classification at -25 dB among three different radio access technologies, namely, DVB-T, WCDMA and IEEE 802.11a, where, consequently, we can reason over the outcome of this classification to differentiate between primary and secondary transmissions. We also show that, through the use of the proposed autoencoder approximate Q-learning technique, such context-aware cognitive radio can achieve better key performance metrics in dynamic spectrum access as compared to previously researched Q-learning algorithms.