Expectation-Driven Debugging: Proactive Deep Learning Model Self-Auditing via Human-in-the-Loop Outlier Analysis

While ProactiV (Prasad et al., 2023) introduced tools to visualize breaking points under input transformations, most current work focuses on static analysis or per-class/instance-level errors. This idea proposes an interactive system where models monitor their own outputs for deviations from human-defined expectations (e.g., "this class should never be predicted for these inputs" or "confidence should not drop below X in such contexts"). When violations occur, the model surfaces not just the outlier instances but also generates counterfactual examples and natural language explanations, inviting domain experts to annotate or adjust expectations in real time. This loop enables rapid identification of both data/model blindspots and domain shifts. Such a system could dramatically improve robustness and trustworthiness, especially in high-risk domains, by bridging the gap between black-box models and expert intuition—a step beyond ProactiV’s focus on visualization and passive analysis.

References:

1. ProactiV: Studying Deep Learning Model Behavior Under Input Transformations. Vidya Prasad, R. J. Sloun, A. Vilanova, Nicola Pezzotti (2023). IEEE Transactions on Visualization and Computer Graphics.