The Goal
A correct image classification does not tell us which parts of an image influenced the model. In a four-person course project at Universitas Gadjah Mada, we explored this question using an ImageNet pretrained ResNet-34 model and sample images.
My Contribution
I worked specifically on the LIME method. I divided images into regions, tested how the model’s predictions changed when those regions were hidden, and visualized which regions mattered most. Our team also presented the project in English.
What We Found
In one rugby ball example, the model identified the object correctly, but the explanation suggested that surrounding details, such as a player’s clothing, also influenced its prediction. That made the result more interesting than the correct label alone: the model appeared to use contextual clues that a viewer might overlook.
Reflection
This project changed how I interpret model performance. A correct prediction is a starting point; explainability methods can help reveal what evidence the model may be using.