It is highly likely that machine learning will play a key role in the next generation technology standard for cellular networks: 5G+ML=6G. In this talk I will review a number of ways I have been involved in exploring and pushing the boundary of that technology, based on a principle which I call Neural Augmentation (NA). In NA we acknowledge that the classical solutions that have been developed in the signal processing community are an extremely strong baseline on which we should build. On the other hand, we know that deep learning is able to learn patterns that are difficult or impossible to detect by humans, if it has access to a sufficiently large dataset and when the domain is narrow enough so that the acquired data can cover it. We argue that by combining classical engineering solutions with deep learning we can learn from smaller datasets and generalize better to out-of-domain situations. In NA we train a neural network to iteratively correct the classical solution. These corrections are hopefully small, and therefore more easy to model. We apply this principle to three problems in wireless communication: error-correction decoding, MIMO demodulation and channel estimation. We find that neural networks are indeed able to improve the state of the art when combined with the classical methods.