Master Thesis

Project Overview

This thesis explores the convergence of Interactive Machine Learning and Neuro-Symbolic Artificial Intelligence. The primary objective is to develop an interactive debugging protocol for hierarchical classifiers. Current methods for explanatory debugging in machine learning are often limited to one-shot interactions, where the machine provides explanations for individual predictions. This project addresses the need for more comprehensive debugging by proposing a multi-round interaction approach. This approach facilitates structured arguments between the machine and the user to validate or challenge predictions and model-related statements.

Objectives

1. Multi-Round Interaction: Develop a multi-round interaction approach to enable more extensive debugging compared to one-shot interactions.

2. Automated Argument Extraction: Automatically extract relevant arguments from the model using eXplainable Artificial Intelligence (XAI) relevance measures.

3. User-Model Interaction: Facilitate dialogical interaction between the user and the machine, allowing structured arguments for debugging.

4. Model Validation: Validate the effectiveness of the interactive debugging protocol on Neuro-Symbolic architectures, including Coherent Hierarchical Multi-Label Classification Networks and Semantic Probabilistic Layers.

5. Feasibility of Bug Correction: Investigate the feasibility of correcting Neuro-Symbolic models with prior knowledge of underlying bugs.

Key Results:

• Correction of confounded model beliefs: this project demonstrates that correcting model beliefs is achievable, especially when prior knowledge of the type of bugs is available.

• Argument selection: This thesis identified metrics for argument selection outperform random selection, reducing interaction time and user annoyance.

• Identifying sources of confounding: The project explores metrics, such as gradients and entropies, for identifying sources of confounding in model predictions.

• Explanatory debugging framework: Combining confounding as a relevance measure and effective debugging, the groundwork for an explanatory debugging process is established.

Future Directions:

• Incorporate Additional Labels: Expand the work by including additional labels that encompass inter-class concepts, allowing for the correction of label-related bugs across the hierarchy.

• Human-in-the-Loop and SPL: Further investigate the relationship between human-in-the-loop techniques and SPL, addressing challenges to enhance model trustworthiness.

• Model Uncertainty Estimation: Explore strategies for accurate model uncertainty estimation, differentiating between aleatoric and epistemic uncertainties.

• User Interaction Evaluation: Assess the overall procedure’s efficacy with real-life users through surveys and questionnaires, gathering qualitative feedback.

Conclusion

This project introduces an innovative interactive debugging protocol that extends beyond one-shot interactions. While serving as a starting point, it has the potential to significantly influence future research in Interactive Machine Learning and Neuro-Symbolic AI.