Target curricula for multi-target classification: the role of internal meta-features in machine teaching

In machine learning, methods inspired by human teaching practices such as the use of curricula, have been fruitful. The bulk of such work has focused on applying a curriculum to the training examples presented to the learning algorithm. However, when there are multiple targets in the learning problem it is also possible to conceive of a curriculum being applied with respect to them. This type of curriculum has received significantly less attention and the key unresolved challenges lay in determining the relative difficulty of learning each target. Logic synthesis is an important aspect of Electronic Design Automation. With the increasing complexity and breadth of designs however comes a significant cost in expert human effort. The most common optimisation-based approaches to automating logic design unfortunately face exponential growth in representation size. Machine learning, which builds models from limited example data, offers a potential remedy to this. The majority of circuit synthesis problems are inherently multi-target making them a good test-bed for target curriculum methods. In this thesis I explore possibilities for target curriculum methods in small-sample machine learning problems using logic synthesis as a test-bed. Using a principle of probably common utility and using intrinsic dimension as a proxy for complexity, I detail two a priori methods, and one self-paced method for generating target-curricula. I also explore a number of ways of applying target curricula within Boolean and neural networks. The results of these studies suggest that target curricula are highly effective in the right circumstances and that using them to directly train digital circuit designs is a compelling direction for future design automation.