Explaining image classifiers with LIME
In the previous section, we have seen how we can easily apply LIME to explain models trained on tabular data. However, the main challenge always comes while explaining complex deep learning models trained on unstructured data such as images. Generally, deep learning models are much more efficient than conventional ML models on image data as these models have the ability to perform auto feature extraction. They can extract complex low-level features such as stripes, edges, contours, corners, and motifs, and even higher-level features such as larger shapes and certain parts of the object. These higher-level features are usually referred to as Regions of Interest (RoI) in the image, or superpixels, as they are collections of pixels of the image that cover a particular area of the image. Now, the low-level features are not human-interpretable, but the high-level features are human-interpretable, as any non-technical end user will relate to the images...