b10-Machine Learning

What is Machine Learning ?

Process of combining inputs to produce useful predictions

How it works

• Train a model with examples(example = input + label)
• Training = adjust model to learn relationship between features and labels
• Feature = input variables
• Inference = apply trained model to unlabeled examples

Learning types

• Supervised learning
  • Regression - Continuous, numeric variables
  • Classification - categorical variables: yes/no
• Unsupervised Learning
  • Clustering - finding pattern
  • No labeled or categorized
• Reinforcement learning
  • Use positive/negative reinforcement to complete a task
    • Complete a maze, learn chess

Neural network

• Neural network - model composed of layers, consisting of neurons
• Neuron - node, combines input values and create one output value
• Feature - input variables used to make predictions
• Hidden layer - set of neurons operating from same input set
• Feature engineering - deciding which features to use in a model
• Epoch - single pass through training dataset
• Deep and Wide in neural network
  • Wide - memorization: many features
  • Deep - generalization: many hidden layers
  • Deep and Wide - both: good for recommendation engines

What is Overfitting?

training model ‘overfitted’ to training data - unable to generalize with new data

Cause of Overfitting

• Not enough training data
• Too many features
• Model fitted to unnecessary features unique to training data: “noise”

Solving of Overfitting

• more data
• make model less complex
• remove “noise”
  • increase “regularization” parameters

AI platform

• Fully managed Tensorflow platform
• Distributed training and predictions
• Hyperparameter tuning with Hypertune

How AI Platform works

• Master - manages other nodes
• Workers - works on portion of training job
• Parameter servers - coordinate shared model state between workers