Landscape

Tools

• Scikit-learn, great entry point
• TensorFlow, more complex, distributed, powerful, by Google

Landscape

Definition

• Program
• Learns from Experience E (training set, training instance or sample)
• With respect to Task T
• And performance measure P
• Improves from experience

Why ML

• Automatically adapting to change
• Deals with complex problem beyond traditional approach
• Help human learn (data mining)

Types

• By human supervision
  • Supervised : labeled data
  • Unsupervised
  • Semisupervised
    • Photo-hosting
  • Reinforcement
    • Agent
    • Observes environment
    • Perform actions
    • Get reward or penalty
    • Learn policy over time
• On the fly
  • Online
    • Out of core learning: learn huge data set in steps
    • Learning rate
      • May learn fast but also forget old experience fast
    • Challenge: garbage in garbage out, so monitor performance closely
  • Batch
• Results
  • Instance based : compare new data points to known data (find value for some close known instance and average)
  • Model based : build a model
    • Model selection
    • Measure model: utility (or fitness) function, how good / cos function : how bad
    • Model parameters (θ)
    • Training: obtain model parameters
    • Inference: make predictions with model

Data terminology

• Label: desired solution in supervised learning
• Features, predictors: values of data

Tasks

• Classification
• Regression: predict a target value
• Clustering
• Anomaly detection
• Visualization & dimensional reduction
  • Feature extraction: Merge several features into one (car mileage and age to wear n tear)
• Associated rule learning
  • Ppl buy a also buys b

What can go wrong

• Data
  • Not enough (unreasonable effectiveness of data)
  • Non-representative data
    • Sampling noise: noise get into
    • Sampling bias: method problem
    • Cleaning data
      • Remove bad instance or fix errors?
      • Some instance miss feature:ignore feature?Ignore instance? Fill in values?
  • Irrelevant features
    • Feature engineering
      • Selection
      • Extraction: combine features
      • Gather new data
  • Over fitting : perform well on training data poor on predictions
    • Reason
      • Model pick up noise
      • Model too complex
    • Fix
      • Choose simpler model
      • Regularization:constraining model
        • Hyper-parameter: parameter to training algorithm
      • More data
      • Reduce noise
  • Under fitting
    • Select more powerful model
    • Feature engineering
    • Reduce constraints

Testing & Validating

• Testing: split data to training and testing set (commonly 80/20)
  • Generalization error (out of sample error): error rate on new cases
  • Over fitting: training error is low, generalization error high
• Validation: if use multiple model and hyper-parameter on same training and testing set to get best model and parameter, it may overfit that set. Keep a further amount of data for validation
  • Cross validation to save data