PhD defence by Erik Arakelyan

Title

Reasoning Inconsistencies and How to Mitigate Them in Deep Learning

Abstract

The recent advancements in Deep Learning (DL) models and techniques have led to significant strides in performance across diverse tasks and modalities. However, while the overall capabilities of models show promising growth, our understanding of their internal reasoning processes remains limited, particularly concerning systematic inconsistencies or errors—patterns of logical or inferential flaws. These inconsistencies may manifest as contradictory outputs, failure to generalize across similar tasks, or erroneous conclusions in specific contexts. Even detecting and measuring such reasoning discrepancies is challenging, as they may arise from opaque internal procedures, biases and imbalances in training data, or the inherent complexity of the task. Without effective methods to detect, measure, and mitigate these errors, there is a significant risk of deploying models that are biased, exploitable, or logically unreliable. This thesis aims to address these issues by producing novel methods for deep learning models that reason over knowledge graphs, natural language, and images. The first part of the thesis contributes two techniques for detecting and explicitly quantifying predictive inconsistencies originating from opaque internal procedures in natural language and image processing models. We systematically evaluate a wide range of model families within novel adversarial setups that explicitly expose those internal procedures, allowing us to quantify significant reasoning discrepancies within these models. To mitigate inconsistencies from biases in training data, this thesis presents a data-efficient sampling method to improve fairness and performance and a synthetic dataset generation approach to rigorously evaluate and enhance reasoning in low-resource scenarios. Finally, the thesis offers two novel techniques to explicitly optimize the models for complex reasoning tasks in natural language and knowledge graphs. These methods directly enhance model performance while allowing for more faithful and interpretable exploration and exploitation during inference. Critically, by addressing reasoning inconsistencies through quantifying and mitigating them with deep learning models, this thesis provides a comprehensive framework to improve the robustness, fairness, and interpretability of deep learning models across diverse tasks and modalities.

For an electronic copy of the thesis, please visit the PhD Programme page.