This is my attempt to (finally!) do some hands-on learning about machine learning. I've learned a bit about the basic theory of ML but I've never tried it myself.

This repo will be mostly copy-pasted code from the TensorFlow Keras tutorials.

To force myself to do some thinking, I'll take notes on what the code is doing. Hopefully this will prevent me from simply blindly copy-pasting.

Theoretically as I continue going through the tutorials I should be able to figure out some of my questions and remove them.

• The clothing classifier is a supervised learning model because we're giving it datasets that are already labeled for training.
• This classifier gives a probability for each possible label, and we pick the highest probability as its prediction.
• A machine learning model is created by connecting several simple layers. Some layers perform simple data manipulation (like Flatten) and others are trained (like Dense, whatever that means).
• "Logits" are (is?) a vector represent the raw predictions generated by a classifier, which are usually then normalized so that humans can understand them.
• Sequential takes in multiple layers and executes them in order.
• The classifier can be very confident and still very wrong.
• loss is the loss on the training dataset; val_loss is the loss on the cross-validation dataset.

• Why do we need to change the image pixel values to be floats from 0 to 1?
  • I tried commenting out those lines and the model became less accurate.
• Does the Dense layer imply the existence of a Sparse layer? If so, why are we using Dense and not Sparse?
• What is an adam optimizer?
• What does relu mean?
• What the heck is a SparseCategorialCrossentropy?
  • This sounds like a randomly generated name that Replit would give you.
• Why does do different runs yield such different predictions?
  • The 13^th image (which is a sneaker but sometimes gets mixed up with a sandal) got a 98% chance of being a sandal on the run after it was classified as a sneaker with around 60% confidence.