Is it a five written in a hurry? Or is it a three with a touch of Asian calligraphy?

Deep models are unsure, as am I. To me it looks like a three, but I can flip a mental switch and it looks like nothing other than a five.

I spent a good month working on the venerable MNIST problem to identify hand-written digits. My model ended up getting 99.82% accuracy on the test set, which I think is really good.

Along the way I became a fan of PyTorch, and switched to it from TensorFlow.

But the stuff I see in the alpha version of TensorFlow 2.0 is pretty neat—in particular the explicit representation of differentiation in tf.GradientTape and the economy of expression afforded by tf.function.

Someday maybe I will be able to write machine learning models directly in NumPy, without the dialectal distractions of PyTorch, and with gradients just where I want them as with TF2. Will it come to pass?

The code for my model, and a runnable Python notebook which demonstrates it, are freely available. Comments are welcome!

PS: My model thinks it is a three also, and it gets marked wrong for that, alas.