Course Notes - Google ML Intro

Course taken in Nov 2018.

Course Details

Title: Introduction to Machine Learning

Link: https://developers.google.com/machine-learning/crash-course/ml-intro

Introduction

Peter Norvig, Director of Research at Google gives the intro explaining how ML helps software engineers - reduce the time programming, customize products for specific people, complete seemingly impossible tasks. Also, philosophical reason - it changes the way you think about a problem. Focus shifts from mathematical science to a natural science, running experiments, use statistics instead of logic to analyze the results.

Tools

Collaboratory - Free Jupiter notebook environment that runs fully in the cloud. https://colab.research.google.com/notebooks/welcome.ipynb

Pandas - Python library for data structures and data analysis. It uses DataFrames and Series. It can load an entire file into a DataFrame, show interesting statistics about data, plot a histogram etc., More info at https://colab.research.google.com/notebooks/mlcc/intro_to_pandas.ipynb. You can run code in this environment and see the results in real-time.

Framing

Supervised ML - System that combines inputs to provide useful predictions on data that was never seen before.

• Label - target that we are predicting; eg: "spam", "not spam". It is the y variable in linear regression.

• Feature - input variable that describes the data (eg: words in the email, to/cc fields, email headers). It is the x variable in simple linear regression.

• Example - A particular instance of data. It can be labeled (includes the feature/s and the label) or unlabeled (contains feature, but no label). We train the models using labeled examples and then use that model to predict the label of unlabeled examples.

• Model - defines the relationship between features and labels. eg: spam detection model may identify some features as strongly associated with spam. Models have two phased - Training (creating/learning the model) and Inference (apply the trained model to unlabeled examples).

• Types of models

  • Regression model: predicts continuous value (price of house in CA)

  • Classification model: predicts discrete values (is the email spam? is this an image of dog/cat/lion)

Descending into ML

• Linear Regression - method of finding a straight line or hyperplane that fits a set of points in the best way. y = wx + b (w is the weight and b is the bias)

• Loss - penalty of bad prediction. Number that indicates how bad the prediction was on a single example.

• Goal of training a model is to find a set of weights and biases that have low loss on average across all examples.

• One of the methods to represent Loss is:

  • L2 Loss (aka Squared Loss): the square of the difference b/w predicted value and observed value of label. MSE = Mean Squared Error; D = data set that contains N pairs of examples (x, y)

$$MSE = \frac{1}{N}\sum_{(x,y) \epsilon D} (y - prediction(x))^2$$