Machine Learning

“Field of study that gives computers the ability to learn without being explicitly programmed” - Arthur Samuel (1959)

• In general the more data the algorithm learns from, the better the result
• Types of ML algorithms based on what problem they solve
  • Supervised learning
  • Unsupervised learning
  • Reinforcement learning
• Hybrid learning problems
  • semi-supervised
    • training data contains very few labeled examples and a large number of unlabeled examples
  • self-supervised
    • framed as a supervised learning problem in order to apply supervised learning algorithms
  • multi-instance
    • individual examples are unlabeled, instead, bags or groups of samples are labeled

Divisions of ml

Supervised learning

Algorithms that learn input x -> output y (label) mappings

• A ml task/function that needs to be provided training data (labeled data (correct answer)) and machine can learn from these results
• You give the learning algorithm examples to learn from that includes the “right” answers
• Notation
  • Training data: Data used to train the model, has both input $x$ and output $y$
    • $x$ (x_train): “input” variable, feature
    • $y$ (y_train): “output” variable, “target” variable (training example targets)
    • $m$: total number of training examples
    • $(x,y)$: single training example (ex. $(2,4)$ means $x=2,y=4$)
    • $(x^{(i)},y^{(i)})$: $i$th training example (x_i, y_i)
  • Model/function
    • $f_{w,b}(x)=wx+b$
  • Estimate/prediction
    • $f_{w,b}(x^{(i)})={\hat{y}}^{(i)}$
  • Parameters: the variables you can adjust during training to improve the model
    • $w$: parameter - weight
    • $b$: parameter - bias
• Real world application
  • First train your model w/ examples input x and the right answers labels y
• Main types of supervised learning
  • Regression
  • Classification

Unsupervised learning

Using a model to describe or extract relationships in data

• A ml task/function that needs NO existing training data
• Data only comes in with inputs x but not output labels y
• Our task is NOT getting an output (ex. identifying if the tumor is benign or malignant)
  • It automatically figures out the STRUCTURE / PATTERN in the data!
• Main types of supervised learning
  • Clustering
  • Anomaly Detection
  • Dimensionality Reduction
  • Association

Reinforcement learning

There is no data, there is an environment and an ML model that generates data and attempts until it reaches a goal

• an agent operates in an environment and must learn to operate using feedback