MNIST (Mixed National Institute of Standards and Technology) is a database of handwritten digits. Compiled by Yann LeCun and colleagues, it’s a classic benchmark problem in machine learning. Pleasingly, you can write a model in Tensorflow that does a decent job of decoding the digits.
Having got to grips with Python and some of the Tensorflow fundamentals, we cannibalised the official Tensorflow tutorial 3 (‘MNIST from scratch’) to make an easy-to-follow tutorial for cracking MNIST using a simple feedforward network. If this is your first time with TF, you might find the Tensorflow Tutorial 1 a better starting point.
We go through the following steps:
- Import the data
- Look at the data
- Figure out normalization
- Compile training, validation, and test sets
- Define model in Tensorflow
- Run the model without training
- Train and test
- Evaluate performance
Download the iPython notebook here.
Any questions or corrections, give me a shout.