Parameters:

• Batch size
• Normalization
• Centering
• Standardization
  *Activation
• Optimizer

Preprocessing: Convert one-hot vector to a pandas Series.

Things to look at:

• Digit value counts
• Heatmaps for each digit
• Correlation for each digit
• VERY simple baseline classifier

As it reads. I want to know exactly how much performance increase I got by spending 4+ hours getting my dependencies just right so I can run TensorFlow on the GPU.

Let's make a benchmark convolutional model and try to beat the previous 92.44% validation accuracy.

Things to review:

• Convolutional Neural Networks
• Keras API

A good way to start is probably to implement simple logistic regression in Keras, and then to move on to CNNs, etc.

E.g., VGGNet or something

Implement logistic regression in Keras.

I want to make sure that we are able to identify and visualize the filters used by the neural network.

To do:

• Install Keras
• Install TensorFlow for the GPU
• Clean up Windows machine environment (keras backend env variable)

We can make the model more robust to translation and rotation by making these transformations ourselves on the MNIST dataset, and using the output as new samples to train on (since we know the correct label).

We need to know how to:

• Translate an image represented as a np array
• Rotate an image represented as a np array

Ideally, we will find or create an API like:

transform_image(rotation=0, translation=(0,0))

Where rotation is the rotation in degrees and the translation is the tuple (dx,dy) .

So far we've only done a baseline CNN. Let's do something more complicated and see how good we can get without extending our data or reaching out to a pretrained network.