How Multi label classification work in supervised learning?

Multi label classification is a deep learning problem solving technique to classify image data that has more than one tagged label (target label). In supervised learning algorithm, the result of multi label classification depends on the target (tagged) label on the images. We may need to classify countries as China, Ethiopia, India and etc, by taking personnel photo as identity. Hence, we simply taking personnel photo and tag (label) the country name with the person photo. Geographical image (map) of the countries is not required.

If we want classify dog type, we take images and tag (label) as bark loud, bark silent, run fast, and run slow, with the corresponding images. So, we may classify the dog as bark loud and run fast. It is a matter of label mapping on the corresponding images. The model can understand both the image feature and corresponding labels. No need of extra images to classify the dog as run fast and bark loud. There is a unique feature for a single doge image. With this unique feature bark loud & run fast labels are mapped.

Consolidate the idea of multi label classification

The following two figure are consolidating these explanations.

The above image includes tree, house, green-land, cloud, and sky. But this image is tag label by only house, tree, and cloud. Because, the researcher needs to classify only with these target labels. Then the supervised model predicts the image only house, tree, and cloud. It cannot classify as both tree, house, green-land, cloud, and sky. Because, even though green-land and sky are parts of the image features, we cannot label them with the image features. This implies that during supervised learning we use image features only as identity for the target label. That means in multi label classification even though labels are not present inside the image, we can tag the target label. From this concept, we can understand labeling is a right of the researchers what they want from the classification result.

The above image is classified as both cat and bird. But there are also trees, wood frame, and electric wire in the image. In supervised learning the machine only understands the tagged (mapped) label with the image feature. The model uses the image features as identity of tagged label. If we label the above image only with trees, wood frame and electric wire, it classifies as these target label. Even though there is cat and bird, the model cannot understand this feature. Because, we label or tag the image only with trees, wood frame, and electric wire. Before reading this blog, I recommend you to see my previous post that talking about CNN. I recommend also to see https://ejssd.astu.edu.et/index.php/EJSSD/article/download/380/89/