Year: 2026
Type: dissertation
Abstract: In evolutionary computing, representation is the encoding we choose for a problem. It defines the search space, determines which operations are possible, and embodies the designer's beliefs about what is important. Generally, it is understood that we should design the representation for the algorithm to efficiently solve the problem, which often coincides with the way humans understand the problem as well. However, this is not necessarily the case. Neural networks are conventionally represented as tensors of weights: ideally suited for gradient descent, but poorly suited for both evolutionary optimization and human comprehension. The work in this thesis spans a period when rapid scaling transformed the landscape of AI, solving many problems using general methods and raw compute rather than specialized representations. This thesis asks: if general methods can solve the problem, is there value to a specialized representation? I argue that the primary value of a specialized representation is not task performance, which scaling of general methods has repeatedly delivered, but human understanding -- and that the primary audience of such a representation is ultimately the human, rather than the algorithm. The chapters begin with neuron-level representations, where I present a permutation-invariant encoding for neural networks which preserves more function under crossover than weight-based representation. Next, I consider subnetworks, treating a pruned subnetwork at initialization as a transferable unit. I show that it transfers across initializations, and propose an evolutionary procedure for re-initializing and re-pruning the subnetwork that improves its performance. Third, I present my contributions to OpenELM, an open-source library for evolution using large language models. Here, the representation is a human-understandable object, such as a prompt or code. Finally, I propose explanation-first XAI, inverting the standard pipeline. Explanations come first and explanation quality becomes the primary target of optimization. I evaluate explanations as communication tools by both modeling and testing audience understanding. Rather than designing the representation for task performance and making explainability a secondary consideration, I put explainability at the forefront, and task performance becomes secondary. I believe this is a choice that we will need to make increasingly often, as models continue to scale while human understanding remains fixed.
Author: Zhou, Ryan
Institution: Queen's University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7220880883
