Abstract
Traditional Explanation-Based Learning (XBL) requires intensive, pixel-level human annotations to guide model focus, creating a significant bottleneck in high-stakes domains like medical imaging. This paper introduces Exemplary Explanation-Based Learning (eXBL), a streamlined approach that refines model behavior using only two "exemplary" explanations—one representing ideal focus and one representing undesirable noise. By leveraging triplet loss, eXBL forces the model to align its internal logic with these global anchors, significantly improving transparency with minimal human effort.
Introduction
In Interactive Machine Learning (IML), the goal is often to ensure that models are not just accurate, but "right for the right reasons." This is particularly critical in fields like healthcare, where a model might achieve high accuracy by latching onto spurious correlations—such as hospital-specific equipment marks in a chest X-ray—rather than actual pathology.
Explanation-Based Learning (XBL) addresses this by incorporating human feedback on model explanations (e.g., saliency maps) into the training loop. However, the standard XBL workflow is prohibitively expensive, requiring experts to provide detailed masks or heatmaps for large portions of the dataset. This paper proposes a more efficient alternative: reducing the human burden from detailed annotation to simple ranking and selection.
The eXBL Framework
The authors introduce Exemplary eXplanation Based Learning (eXBL), which simplifies the feedback loop by identifying two global anchors:
- $C_{good}$ (The Gold Standard): An explanation where the model correctly focuses on the relevant biological or diagnostic features.
- $C_{bad}$ (The Cautionary Tale): An explanation where the model focuses on background noise, artifacts, or irrelevant regions.
Optimization via Triplet Loss
To integrate these exemplars, the framework utilizes a Triplet Loss function, a technique traditionally used in metric learning and face recognition. In the eXBL context, the triplet is defined as:
- The Anchor: The GradCAM explanation generated for the current training image.
- The Positive: The "Good" exemplary explanation ($C_{good}$).
- The Negative: The "Bad" exemplary explanation ($C_{bad}$).
The loss function penalizes the model if its current explanation is mathematically closer to the "bad" example than the "good" one. By minimizing this distance, the model is incentivized to produce explanations that mirror the high-quality exemplar across the entire dataset, effectively propagating the logic of just two images to thousands of others.
Experimental Setup
The study evaluated the eXBL approach using the Covid-19 Radiography Database, which contains chest X-ray images across four categories: Covid, Normal, Lung Opacity, and Viral Pneumonia.
- Model: MobileNetV2.
- Data Split: 3,200 training images, 1,200 validation images, and 800 test images.
- Ground Truth: The authors utilized existing radiologist-provided feature masks to objectively select the "good" and "bad" exemplars, simulating the selection an expert would make in a real-world scenario.
Results and Discussion
The comparison between an unrefined MobileNetV2 and the eXBL-refined version revealed a compelling trade-off between transparency and raw performance.
Improved Interpretability
The eXBL model demonstrated a measurable improvement in Activation Recall (+0.02, or +3%). This metric indicates that the model's heatmaps overlapped more accurately with the actual lung regions identified by medical experts. Qualitatively, visual inspections confirmed that while the baseline model often focused on image edges or random noise, the eXBL model consistently centered its attention on the relevant lung anatomy.
The Accuracy-Interpretability Trade-off
Refining the model's focus led to a slight decrease in classification performance, with accuracy dropping from 95% to 91%. This highlights a classic tension in AI: when a model is forced to ignore "shortcuts" (like specific equipment marks that might be predictive in a biased dataset but are medically irrelevant), its performance on that specific dataset may decrease. However, the resulting model is arguably more robust and trustworthy for clinical deployment.
Conclusion and Future Directions
The eXBL method demonstrates that just two exemplary images can effectively steer the explanation logic of a complex model. This significantly lowers the barrier to entry for expert-in-the-loop refinement.
Future work will focus on:
- Active Learning: Developing methods to help experts identify the most "informative" explanations to use as exemplars.
- Scalability: Investigating whether a small "budget" of exemplars (e.g., 5-10) provides diminishing returns or significantly better alignment.
- Mitigating Performance Loss: Exploring whether the 4% drop in accuracy is a result of removing bias or if the exemplary constraints are overly restrictive, potentially masking subtle but valid patterns in the data.
Summary updated