Week 1: Introduction to Machine Learning

Supervised Learning

• The most common types of algorithms being used today
  • learns from 'input' to 'output label'
  • use cases: spam filtering, speech recognition, machine translation, online advertising, self driving
• Gives the algorithm datasets to learn from labelled "correct answers"
  • Regression modeling is to predict a range of numbers
  • Classification modeling is to predict a limited set of class/category

Unsupervised Learning

• Learns from unlabelled output data to find patterns or interesting properties
  • learns from 'input' only as there are no 'output label'
• Anomaly detection
• Clustering
• Dimensionality reduction

[[Linear Regression Model]]

• Fitting a straight line to a dataset
  • produce a function that produces a y-hat (estimated y)
  • formulated as a wx+b since it is a linear function
  • to train/adjust the unknown parameters/coefficients/weights of the function we need to construct a cost function
• Cost function is a way to quantitatively evaluate the goodness of fit of the function
  • $J(w, b) = rmse/mse$
  • Can be minimized with [[gradient descent]] to find the optimal parameters

Week 2: Regression with multiple input variables

Multiple Linear Regression

• Using multiple features to estimate one response variable
  • Different from multivariate regression which estimates multiple response variable
• Vectorization with numpy is used to speed up the algorithm of the dot product between weights and features
  • This is done by doing the operation in parallel
• Gradient descent works similarly but instead of taking the derivative of the cost function with respect to one feature's weight and updating one weight -> you update all the weights by taking partial derivatives of the cost function with each weight
• Just a note that an alternative solution to the linear regression is the [[normal equation]]
  • It only works for linear regression because it is a closed form solution to this
  • So it is not generalized solution (unlike gradient descent) and it could be slower if the features are much higher

Gradient descent in practice

• To make gradient descent more efficient, one should feature scale with some [[Feature Scaling Techniques]]
  • This will make the cost function contour to be evenly spaced and not be susceptible to jumping

Week 3

#machine-learning #course