Reading Notes of Grokking Deep Learning Chapter 4

Foreword

Following the last post, I will continue to summarize key points of Chapter 4 in this blog.

Nearly all supervised learning projects can be divided into 3 parts,

predict, compare, learn

so chapter 3 focus on predicting part and chapter will focus on comparing and learning.

Chapter 4

Simple Code and Simple Math but the concept is very important:

Learning in brief

Learning is adjusting the weight to reduce the error to 0!

With derivatives, you can pick any two variables in any formula, and know how they interact! It is the sensitivity between two variables.

Why learning rate alpha

the simplest way to prevent overcorrecting weight updates!

Not solving the overshooting problem just a way to make weight update smaller.

Transfer Learning Benefits

The benefits of transfer learning:

  • Transfer learning needs less training data
  • Transfer learning makes the learned model generalizes better
  • Transfer learning makes training process less brittle
  • Transfer learning makes deep learning easier
Written on April 21, 2019