Beautiful visual model interpretation of classification strategies— Kannada MNIST Digits Recognition

The Kannada MNIST dataset is a great recent work (details here), and I’m delighted that it is available to the public as well. I’m sure pretty soon the community here would be posting state of the art accuracy numbers on this dataset. Which is why, I’m doing something different.

Instead, we will try to visualize, try to see what the model sees, assess things pixel by pixel. Our goal would be interpretability. I’ll start with the ‘simplest’, easiest to interpret algorithm in this article. Hopefully, I’ll post results with other modeling techniques in later article.

To reiterate and clarify: I…