This project uses ResNet-50 to classify images and explain the predictions made by the model. The explanations are generated using two popular model explanation techniques: LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations) with the partition explainer. These methods help to understand the reasons behind the model's classification decisions by highlighting important features in the images.
- LIME: Generates local explanations for individual predictions by perturbing the input and observing changes in the model's output.
- SHAP (Partition Explainer): Uses a game theory approach to distribute the importance of features among different parts of the input, making it easier to understand the global impact of features on predictions.
Given Inputs:
| Input 1 | Input 2 |
|---|---|
 |
 |
The model shows that the head and neck of the bird were the most important parts of the image that positively guided the model to predict the image as American_egret class. However, parts of the background have negatively affected its decision.
For the second image, certain parts of the boat and water contributed positively to the classification as a lifeboat, while other parts had a negative impact. Notably, the lifeboat class was the model’s fourth prediction, with a confidence of 0.039%. This suggests that the large white section of the boat which resembles a real lifeboat, may have contributed to this prediction.
As shown in the figure, in the first image, regions including parts of the neck and the back of the bird have positively contributed to the correct classification as American_egret, while the bird's eyes and beak have undermined the possibility of the first class and strengthened the likelihood of belonging to other classes like crane or little_blue_heron.
For the second image, most parts of the boat, in addition to part of the background, led to the decision to classify this input as a speedboat. However, similar to the LIME result, big white parts of the boat led the model to classify it in other classes like lifeboat.