跳转至

Lecture 1 2022.2.22

What is Artificial Intelligence？

Brief History

总共有三种流派，反映了人们对于人工智能的艰辛的探索。也表现的是人们并不完全掌握对于“智能”的理解。

为什么会有人工智能的寒冬?

早期 AI

  1. 1952 lookahead search
  2. 1956 search tree & heuristics search

→ 通用人工智能的乐观氛围

1956, Dartmouth

十个人，开了两个月的会，提出了一个新的词：Artificial Intelligence

Every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be make to simulate it.

First AI Winter

1966, ALPAC report cut off government funding for MT(machine learning)

Problems:

  1. Limited computation: exponentially complexity in our method
  2. Limited information: the problem of real world is too difficult(number of words, objects, information ...)

by-product: lisp & lisp machine

Knowledge-based Systems

Inspiration from Human Brain: learn from what you know

Giving rules -- from human

Examples:

  • DENDRAL(1960s): rules for guessing the property of unknown organic molecules

  • MYCIN(1970s): ～600 rules from human doctors; only can diagnose blood infections

→ problems: not general

Expert System

Problems:

  1. Deterministic rules could not handle the uncertainty of the real world
  2. Rules can get too complex to create and maintain.

Second AI Winter

Neural Networks

simulate humans neural

Minsky: cannot handle xor!

Convolutional Neural Network

1980, 一个神经元只能看到一部分(Receptive field)

1986, backpropagation(反向传播) algorithm is popular, training multi-layer networks

→ break the limitation of linear classifier

Application: 1989, Yann LeCun applied CNN to recognizing handwritten digits

---> 从造人变成了造工具，可以认为是一个理念的转变

Deep Learning

AlexNet(2012): deep convolutional networks on ImageNet powered by GPU

AlphaGo(2016): deep reinforcement learning, MCTS

Bayesian Networks

1985, use graph model for reasoning under uncertainty

实际问题变成随机变量之间的相关和因果关系；

point -- things

edge -- rule

Support Vector Machines

1995 maximize margin

总结：什么样的算法能火起来呢？刻上了时代的烙印...

什么样的需求？什么样的条件？

Overview of AI knowledge structure

Two Views of AI

What are the two views?

  • Intelligent Agent: think & act like a human
  • six aspects: perception（感知）, robotics（执行）, language（语言）, knowledge（知识）, reasoning, learning（学习）
  • AI Tools: Do something that human cannot do

Intelligence Agent

observe environment, perform activity

in six aspects

machine can only do narrow tasks, with millions of examples

AI Tools

planning, scheduling, pattern recognition, web mining, prediction

We need to be rational: it is a better tool, but not a perfect tool

AI + SCIENCE

e.g. GPT-3, 175 billion, alike to neural cell's number in human mind

Language Models are Few-Shot Learner

尤其是在中国大陆，最不健康的一股思潮：我的模型比你的大，我的参数比你的多！

500-word article writing: 看起来很好，似是而非，缺少精确性

e.g. BigGAN 无中生有的生成的图片，不受控制，不能精细化生产

e.g. MuZero, DeepMind 的下棋的问题，less prior knowledge， more tasks

Question

  • Interpretation
  • Security
  • Bias

Course Overview

学习 AI 的精华

THIS IS A DEMANDING COURSE.