Deep Learning 2 - Your First Neural Network

MNIST, architecture, compiling, data prep, training, predicting

One of the best ways to learn neural nets is by doing. So, in this case, we’ll grab a classical dataset, and crack open a neural net, and explain what’s going on.

For this post, you’ll want to ahead, and download the tensorflow, and keras libraries to your python.

1 - MNIST Dataset

Let’s say we want to come up with an algorithm that can analyze handwritten digits, and classify them from 0 to 9. To do this, we’ll need several photos of handwritten digits, and we’ll need them classified for us, so that we can pass the X data (the photos), and the target Y data (classification) to our neural net to see the pattern.

Luckily, this data already exists, it’s called the MNIST digits data. To load this in Python, simply run the following code:

from tensorflow.keras.datasets import mnist
(train_images, train_labels), (test_images, test_labels) = mnist.load_data()

Once you have these pictures loaded, you’ll see that you get a numpy array, where train_images.shape has the following dimensions (60000, 28, 28). This is telling us that our training images data has 60,000 photos, each photo having the length & width of 28.

You can use basic indexing, and matplotlib to plot the images and view them, this is what they look like:

If you are like me, and use Pycharm, you can just click on the variable, and have it load the first one for you. Feel free to guess which number this is:

2 - Network architecture