1. This blog posts aims at instilling an understanding of "Label Smoothing" in the minds of reader. 2. At the end of post, it is concluded that if your dataset has symantically similar classes or mislabelled images, label smoothing can improve your model accuracy. 3. This blog also explains a new way of thinking about increasing or decreasing of logits which was mentioned in the paper "When Does Label Smoothing work" by Rafael Müller et.al
Don't forget to tag @Abhimanyu08 in your comment.