This thesis explores how different conceptualisations of task relationships may inform transfer in the context of cognitive training. There are three large empirical chapters: The first uses a novel analysis pipeline applied to a composite of pre-existing datasets, to explore training outcomes across four popular tasks. Specifically, I used two unsupervised machine learning algorithms to identify multivariate task profiles and sub-groups; I then used these to ask whether, and how, task profiles change following training, both across and within sub-groups. The second empirical chapter presents an online cognitive training study exploring transfer patterns within a set of bespoke tasks that are nested hierarchically, and systematically related, according to simple task features. This approach revealed that training at the top of the hierarchy can yield benefits that cascade to lower-level components, but not the reverse. Furthermore, I quantified the overlap between tasks in different ways and then tested which metric best predicts patterns of transfer. In this case, the presence of one particular shared feature across tasks was the best predictor of transfer. In the third and final empirical chapter another large-scale online training study focussed on a set of change detection tasks, again with a nested set of interrelationships. Again, the results speak to the feature specific nature of transfer patterns, but also show that specificity is context dependent. In this case transfer could be bidirectional within the hierarchy. That is, training lower-level components would yield benefits for those same components and in some cases across components, where they appeared within more complex tasks, and likewise, training the complex task would yield benefits in lower-level constituent components. In the final discussion I integrate across these empirical findings, consider how these results fit within the broader cognitive literature and theories of transfer, along with some recommendations for future research.