Programming-by-Example, and code synthesis in general, is a field with many different sub-fields, involving many forms of machine learning and computational logic. With advantages and disadvantages to each, attempts to build effective hybrid solutions would seem to be a promising direction. Transfer Learning (TL) provides a good framework for this, as it allows one of the classic code synthesis techniques, Genetic Programming, to be augmented by past success, to target a particular code synthesis system to the problem domain it is facing. TL allows one type of machine learning algorithm, in this thesis a neural network, to support the core GP process, and combine the strengths of both. This thesis explores the concept of hybrid code synthesis approaches, and then brings the identified strongest elements of each approach together into a single neural network driven Transfer Learning system for Genetic Programming. The TL system operates autonomously, without any human intervention required after the problem set (in example only format) is presented to the system. The thesis first studies how to structure a training corpus for a neural network, across two different experiments, exploring how the constraints placed on a corpus can result in superior training. After this, it studies how GP processes can be guided, to ensure that a hypothetical NN guide would be useful if it could be created and how it can best assist the GP. Finally, it combines the previous studies together into the full end-to-end TL system and tests its performance across two separate problem domains