Thesis Defence: Structured Representations for Interpretable AI: Knowledge Graphs and Latent Trees

Ethan Thoma, supervised by Dr. Gema Rodriguez-Perez, will defend their thesis titled “Structured Representations for Interpretable AI: Knowledge Graphs and Latent Trees” in partial fulfillment of the requirements for the degree of Master of Science in Computer Science.

An abstract for Ethan Thoma’s thesis is included below.

Defences are open to all members of the campus community as well as the general public. Registration is not required for in-person defences.

---

Abstract

Opaque AI systems limit deployment in safety-critical applications. This thesis investigates whether forcing models to use structured representations can improve interpretability without sacrificing performance through two case studies: knowledge graphs for bug triage and tree-structured reasoning for mathematical problem solving.

We convert GitHub issues into typed knowledge graphs using REBEL-LARGE relation extraction and BERT embeddings. Human evaluation shows 48% of the graphs miss nuanced information, particularly implicit developer consensus and multi-turn argumentation. Despite these gaps, classifiers trained on graph features achieve macro-F1 of 0.4762, beating text-only baselines at 0.3564 and enabling maintainers to inspect which entities drive triage decisions.

Dynamic Compute Tree Modelling (DCTM) injects 32 parallel latent nodes at layer six of GPT-2, learning tree structure through Gumbel-annealed parent prediction. Operation-conditioned adapters prove essential, achieving 20.7% validation loss improvement over baseline on mixed reasoning tasks. Mid-layer injection works best, consistent with findings that middle layers contain abstract task-agnostic representations. Systematic generalization experiments show 12-20 point improvements on out-of-distribution tests, though 32-node capacity constrains performance on complex problems. Interpretability analysis reveals moderate operation alignment (55-65%), with trees providing coarse-grained debugging signals through structural consistency metrics.

These findings challenge the assumption that interpretability requires perfect structural fidelity. Graphs with 48% coverage gaps improved classification by 33.6% and tree structures with partial operation alignment maintained advantages on out-of-distribution tests. This work establishes design principles for structured interventions, including mid-layer composition, sparsity as an interpretability prior, and task-appropriate representation choice, offering pathways toward AI systems that are both powerful and transparent.