Lesson Three

Precompute is only a way to cache some steps to speed it up
Study logs and Exponentiation math
Softmax is x/sum
- sum of softmax should be 1
- inputs should be >= 0
Since images are matricies, multiply by some number to make it less washed out
Fast AI you can resize all the images with a data.resize() to speed up training

bn_freeze

If you're using a deeper model, if the dataset is similar to the dataset that trained the model, this will help (more about this later in the course)
- Stands for batch normalization

predict_array

You can use this to predict a single image while passing in your image
- If you index an array with [None] it turns the array input into a tensor since that's what predict is expecting
You're have to transform your image using tfms_from_model

Intro to theory behind Convolutional Neural Networks

Matrix of N x N is multiplied by each N x N section of the image and outputs the next layer (Excel Example)
This is done again and again until the last layer
Some layers may check for left edges, lower edges, eyes, etc

Multi-Label Classification

Instead of softmax use the sigmoid function

Visualize model

learn.summary
- Shows what each layer is doing