This example demonstrates the use of TensorFlow and neural networks to classify handwritten digits.
Simply write a number in the box below and hit "Guess".
In order to simulate the data on which the model was trained, we use an image size of 28 by 28 pixels, scaled much larger, creating a blurring effect.
The accuracy of the model is actually around 99.2%, which is pretty good. It doesn't appear accurate in some cases because the input data is not quite like the data it was trained on. The data is dissimilar due to the way that it's processed, and the fact that numbers drawn on a computer screen do not look the same as numbers which are handwritten.
The model's accuracy is around 99.2%. There are many ways of improving this, including applying affine transformations to create new training images, and using a committee of many networks.
The predictions actually take place in less than a millisecond. The time you see is the time to load the model into memory from disk. This could be sped up by loading the model once into memory and keeping it there.
It only takes a day or two to create a model which can recognize digits. This is due to an abundance of training data and previous work.
Machine learning can do complex image recognition, classification, and regression tasks. There are many examples on the internet of these types of models and their results.>