Transparent and Coherent Procedural Mistake Detection

Abstract: Procedural mistake detection (PMD) is a challenging problem of classifying whether a human user (observed through egocentric video) has successfully executed a task (specified by a procedural text). Despite significant recent efforts, machine performance in the wild remains nonviable, and the reasoning processes underlying this performance are opaque. As such, we extend PMD to require generating visual self-dialog rationales to inform decisions. Given the impressive, mature image understanding capabilities observed in recent vision-and-LLMs (VLMs), we curate a suitable benchmark dataset for PMD based on individual frames. As our reformulation enables unprecedented transparency, we leverage a natural language inference (NLI) model to formulate two automated metrics for the coherence of generated rationales. We establish baselines for this reframed task, showing that while VLMs struggle off-the-shelf, their accuracy, coherence, and efficiency can be improved by incorporating these metrics into common inference and fine-tuning methods- though not without tradeoff. Lastly, our multi-faceted metrics visualize common outcomes, highlighting areas for further improvement.

• The paper introduces a self-dialog framework that reformulates procedural mistake detection as an interpretative task, enhancing AI transparency.
• It integrates vision-and-language models with coherence metrics to improve explanation relevance, informativeness, and detection accuracy.
• Fine-tuning with coherence-based re-ranking and in-context learning significantly boosts model responsiveness for procedural error detection.

Explainable Procedural Mistake Detection

The paper "Explainable Procedural Mistake Detection" addresses a critical problem in the context of automated task guidance: the challenge of procedural mistake detection (PMD) through the observation of human actions via egocentric videos. This research expands upon existing methodologies by introducing an innovative approach to enhance visibility into the reasoning processes of AI models used in PMD. It focuses on reformulating PMD as an explanatory task through a carefully designed self-dialog framework.

Problem Setup and Methodology

The primary aim of the paper is to tackle PMD by reconceptualizing it as an interpretative task where the detection of procedural errors is supported by an accompanying dialog. This dialog includes a series of generated questions and answers intended to substantiate the decision-making process. The reformulation seeks to advance transparency and coherence in machine interpretation by leveraging vision-and-LLMs (VLMs) combined with a natural language inference model. Crucially, the authors propose two automated metrics for evaluating the coherence of explanations: relevance and informativeness.

The methodology involves using a fine-tuned VLM to generate questions and answers iteratively, thus providing insights into the reasons behind procedural classifications. To validate and improve this process, the study applies coherence metrics during question selection and generation phases. Different architectures are tested, including LLaVA, InstructBLIP, and Llama 3, alongside refined interventions such as coherence-based re-ranking of question candidates and fine-tuning procedures that employ preference optimization.

Key Findings

Experimental results are obtained through a specially curated dataset (Ego4D-PMD), which includes annotated video frames representing various procedural tasks. The study reports several significant findings:

1. Coherence Metrics: Inclusion of coherence metrics notably enhances the relevance and informativeness of generated explanations. This improvement is manifested in higher PMD accuracy while also streamlining self-dialog length and converging more efficiently toward decisions.
2. Model Adaptations: It is demonstrated that coherence-based question re-ranking and additional in-context learning substantially ameliorate both the relevance and quality of VLM's output for PMD, supporting more accurate mistake detection than baseline models.
3. Fine-Tuning Impact: Fine-tuning strategies that incorporate coherence feedback show potential benefits in model responsiveness and efficiency, indicating that specialized training further strengthens task performance.

Practical and Theoretical Implications

The study's implications are pronounced in both practical and theoretical domains. Practically, the proposed framework extends the boundaries of automated task systems by integrating explainability, which is paramount for real-world applications where transparency and user trust are critical. Theoretically, it introduces a novel approach to the challenge of aligning VLMs' interpretative capabilities with human reasoning by bridging perceptual inputs with semantic content through structured reasoning.

Moreover, this research sets a precedent for future explorations in AI, suggesting a shift toward explainable and interactive AI systems. The integration of advanced coherence metrics could potentially be transferred to other AI domains that demand transparency and accountability, such as healthcare or autonomous systems.

Future Directions

Given the foundation laid by this research, future endeavors could focus on further refining the dialog structure and exploring scalability in diverse procedural contexts. Additionally, as VLM technologies evolve, integrating richer modalities, such as temporal reasoning across video sequences, could offer deeper insights into the holistic understanding of procedural tasks.

Overall, the reformulation of PMD into an explanatory task as proposed in this study presents an important step forward in achieving a more interpretable and accessible AI that mirrors human-like understanding and decision-making fidelity.