Exploration 1.0: Introduction

Machine Learning is Used Everywhere

• Today, ML is used everywhere in your daily life, even though you might not notice it. The human civilization will inevitably depend more and more on ML.

• ML is a subfield of AI, along with other subfields such as natural language processing (NLP), computer vision (CV), AI search, information retrieval, robotics, etc. Nowadays ML is the hottest subfield in AI because all other subfields make heavy use of ML.

• Within ML, there are many subsubfields, such as deep learning (DL), reinforcement learning (RL), unsupervised learning, structured prediction, etc.

• AI Landmarks: more and more ML

  • IBM Deep Blue (1997): computer chess. AI search; no ML.
  • IBM Watson (2011): question answering. Mostly classical NLP, with very little ML.
  • Google DeepMind AlphaGo (2016): computer Go. A lot of ML: Deep reinforcement learning + AI search.
  • OpenAI ChatGPT (2022). All ML: deep learning + reinforcement learning.

What is Machine Learning

• Programming = Automation

  • caveat: you have to write the programs yourself

• Machine Learning \(\simeq\) Automating programming = Automation\(^2\)

  • Think about translation: how to write a program to translate from English to Chinese?
  • Well, you can write LOTS AND LOTS of rules, including all dictionary mappings such as “bank” => “银行 (yinghang)”, and phrasal translations such as “the X of Y” => “Y 的 (de) X”.
  • This is extremely tedious and laborious, and won’t work very well at all: How to choose the best rule in case of ambiguity (“bank” => “银行” vs. “bank” => “河岸”)? How about new words or expressions that keep emerging? <ost importantly, how to deal with another language pair? Hire a bunch of bilingual experts to write a new set of rules??? But there are so many language pairs in the world and they keep changing!
  • This is exactly the reason why “rule-based AI”, the dominating AI paradigm before ML, fell out of fashion in the 1990s. It is simply too costly, not working well, and not scalable.
  • ML systems, by contrast, do not ask humans to write rules. They learn rules and their weights (scores) directly from data. In the case of translation, this data is “parallel text” (“bilingual text”, or “bitext”) which is a collection of aligned sentence pairs from two languages. For example, take a user manual from any electronic product; it must be written in parallel text of many languages (this is beyond “bitext”). Let’s say we have 1 million <English, Chinese> sentence pairs. We train an ML system to extract rules automatically from this data set, just by co-occurrence statistics. If the word “bank” tends to co-occur with “银行 (yinghang)” in many sentence pairs, then the system guesses that they are translations of each other. Gradually the system can figure out more mappings, including phrasal ones like “the X of Y” <=> “Y 的 (de) X”. So now you’re relieved of the tedious job of writing rules.
  • But there is a caveat: you have to write the ML program (which is highly non-trivial!). Well, the good thing is, once you wrote such a program, and it works well on one language pair, you can apply it to many other language pairs automatically. Well, this was how Google Translate worked for many years, before switching to deep learning-based neural machine translation around 2016.

• Deep Learning \(\simeq\) Automating machine learning = Automation\(^3\)

Course Roadmap

• Unit 1: Intro, \(k\)-NN (weeks 0-3)

  • Quiz 1 (ML basics) and Quiz 2 (linear algebra review)
  • HW1: income classification with \(k\)-NN

• Unit 2: Linear Classification and Perceptron (weeks 4-5)

  • HW2: sentiment analysis with perceptron

• Unit 3: Regression (weeks 6-7)

  • HW3: housing price prediction (Kaggle)

• Unit 4: A Taste of Deep Learning (weeks 8-9)

  • HW4: redo HW2 with word2vec embeddings

• Unit 5: Paper Review (week 10)

  • HW5: critic of cutting-edge ML research

Videos

• Intro, part 1
• Intro, part 2