Stanford Pratical Machine Learning-机器学习简介

这一章主要简单介绍一下机器学习，包括机器学习的类型之类的。

Types of ML Algorithms

• Supervised：Train on labeled data to predict labels

  • Self-supervised: Labels are generated from data. (E.g. word2vec, Bert)

• Semi-supervised：Train on both labeled and unlabeled data, use models to infer labels for unlabeled data

  • E.g. self-training

• Unsupervised：Train on unlabeled data

  • E.g. clustering, density estimation(GAN)

• Reinforcement learning：Use observations from the interaction with the environment to take actions to maximize reward

• Tips:

  • We can design supervised training tasks for unlabeled data:

    • Self-supervised learning: generate labels from data. E.g. word2vec, BERT
    • GAN: generating fake data with trivial label from unlabeled data

  • Training tasks can be different from how the model is evaluated / used.

Components in Supervised Training

• Model

  • A parameterized function to map inputs to label
    • Model parameters VS hyper parameters
    • E.g. listing house -> sale price

• Loss

  • The measure of how good the model does in terms of predicting the outcome
    • E.g. classification / regression / contrastive / triplet / ranking
    • E.g. $(predict_price - sale_price) ^ 2$

• Objective

  • The goal to optimize model params for
    • E.g. minimize the sum of losses over examples

• Optimization

  • The algorithm for solving the objective

Types of Supervised Models

• Decision trees: Use trees to make decisions
• Linear methods: Decision is made from a linear combination of input features
• Kernel machines: Use kernel functions to compute feature similarities
• Neural Networks: Use neural networks to learn feature representations

Summary

References

1. slides