๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ

๐“ก๐“ธ๐“ธ๐“ถ5: ๐’ฆ๐‘œ๐“‡๐‘’๐’ถ ๐’ฐ๐“ƒ๐’พ๐“‹/๊ฒฝ์˜์ •๋ณด์‹œ์Šคํ…œ(BUSS215)

[๊ฒฝ์˜์ •๋ณด์‹œ์Šคํ…œ] 1. Artificial Intelligence

* Exemplary System์€ ์‹œํ—˜๋ฒ”์œ„์— ํฌํ•จ๋˜์ง€ ์•Š์œผ๋‹ˆ ์ƒ๋žตํ•œ๋‹ค

 

Deductive : ์—ฐ์—ญ๋ฒ• (Idea -> Observations -> Conclusion)

-> Rule-Based

Inductive : ๊ท€๋‚ฉ๋ฒ•

-> Case-Based (Observations -> Analysis -> Theory)

 

Artificial Intelligence (AI)

A huge set of tools for making computers behave intelligently

Expert system : ์‚ฌ๋žŒ์ด ์•„๋Š” ๊ฒƒ์„ ๊ธฐ๊ณ„์—๊ฒŒ ๊ทธ๋Œ€๋กœ ์ฃผ์ž…์‹œ์ผœ ์˜์‚ฌ๋‚˜ ๋ณ€ํ˜ธ์‚ฌ์ฒ˜๋Ÿผ ์ถ”๋ก ํ•˜๋Š” ๊ธฐ๊ณ„๋ฅผ ๋งŒ๋“ค ์ˆ˜ ์žˆ๋‹ค๋Š” ์ƒ๊ฐ์—์„œ ๋‚˜์˜จ ๊ฒƒ์œผ๋กœ Rule based์ธ AI ์ข…๋ฅ˜ ์ค‘ ํ•˜๋‚˜. ํ•˜์ง€๋งŒ ์ด๊ฒƒ์€ ์ธ๊ฐ„์˜ ๋Šฅ๋ ฅ๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ ํ•œ์ฐธ ๋ถ€์กฑํ•จ

  • ๋‹จ์ง€ ์ธ๊ฐ„์„ ๋ชจ๋ฐฉํ•œ ๊ฒƒ์ด๊ณ 
  • ์ธ๊ฐ„์ด ์ƒ๊ฐํ•˜๊ณ  ์ธ์ง€ํ•˜๋Š” ๋ณต์žกํ•œ ๊ณผ์ •์€ ๋ฐ˜์˜ํ•˜๊ณ  ์žˆ์ง€ ์•Š์Œ
  • ์•„๋ฌด๋ฆฌ ์—ด์‹ฌํžˆ ์—ฐ๊ตฌํ•ด๋„, ์šฐ๋ฆฌ๊ฐ€ ์•„๋Š” ๊ฒƒ์ด ๋„ˆ๋ฌด ์ ๊ณ 
  • ์ธ๊ฐ„์ด ์•„๋Š” ์ง€์‹์„ ์ถ”์ถœํ•ด์„œ ๊ธฐ๊ณ„์— Rule์„ ์ž…๋ ฅํ•ด์•ผ ํ•จ.
  • ๋ˆ๊ณผ ์‹œ๊ฐ„์ด ๋“ค๊ณ , ์ƒˆ๋กœ์šด ์ง€์‹์„ ์•Œ๊ฒŒ ๋  ๊ฒฝ์šฐ rule ์—…๋ฐ์ดํŠธํ•˜๋Š”๋ฐ endless loop ๋Ž

๊ทธ๋ž˜์„œ ๊ธฐ๊ณ„๊ฐ€ ์Šค์Šค๋กœ ํ•™์Šตํ•˜๋Š” Machine Learning(ML)์ด ๋“ฑ์žฅํ•˜๊ฒŒ ๋จ!

 

Machine Learning

A set of tools for making inferences and predictions from data

 

๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ real-world process์˜ ํ†ต๊ณ„์  ํ‘œํ˜„์ž„!

New input -> [Model] -> Outcome

 

๋ฌด์—‡์„ ํ•˜๋Š๋ƒ!

1. ๋ฏธ๋ž˜์˜ ์ผ์„ ์˜ˆ์ธกํ•œ๋‹ค

2. ์–ด๋–ค ์ผ์ด๋‚˜ ํ–‰๋™์˜ ์›์ธ์„ ๋ถ„์„ํ•œ๋‹ค

3. ํŒจํ„ด์„ ์ถ”๋ก ํ•œ๋‹ค

 

 

ML์˜ ์ข…๋ฅ˜

 

 

 

Supervised

: labeled data๋ฅผ ํ•™์Šตํ•œ ๋’ค, unlabeled ๋œ ์ƒˆ๋กœ์šด ์ •๋ณด๋ฅผ ์˜ˆ์ธก

- Classification : item์„ ์นดํ…Œ๊ณ ๋ฆฌ๋กœ ๋ถ„๋ฅ˜/์˜ˆ์ธก

- Regression : ์ˆซ์ž ์˜ˆ์ธก

 

Unsupervised

: unlabeled ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•˜์—ฌ output ์˜ˆ์ธก

- Clustering : ๋น„์Šทํ•œ item๋ผ๋ฆฌ ๋ถ„๋ฅ˜ํ•˜๊ธฐ

- Dimension Reduction(Generalization) : ์ฃผ์–ด์ง„ ๋ฐ์ดํ„ฐ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒˆ๋กœ์šด ๊ฒฐ๊ณผ๋ฌผ ์ƒ์„ฑ

- Association : ์—ฐ๊ด€๊ด€๊ณ„, ์ธ๊ณผ๊ด€๊ณ„ ์˜ˆ์ธก

 

 

Data set์˜ ์ผ๋ถ€๋Š” Trainํ•˜๋Š”๋ฐ, ์ผ๋ถ€๋Š” Testํ•˜๋Š”๋ฐ ์‚ฌ์šฉํ•จ (80:20์˜ ๋น„์œจ)

X1~X5 -> Features

Y -> Target Variable 

 

Unsupervised learning์€ Training data์— target variable ์—†์ด feature๋งŒ ์กด์žฌํ•จ

 

 

 

ํ•˜์ง€๋งŒ ML ์—ญ์‹œ ๋“ฑ์žฅํ•œ ์ดํ›„ ํ•˜๋ฝ์„ธ๋ฅผ ๊ฑท๋Š”๋ฐ... ๊ทธ๋Ÿฌ๋˜ ์ค‘ neural net์˜ ๋“ฑ์žฅ์œผ๋กœ '๊ฐœ๊ฐ™์ด ๋ถ€ํ™œ'
๊ทธ๋ฆฌ๊ณ  deep learning์˜ ์„ธ์ƒ์ด ์—ด๋ฆฐ๋‹ค

 

Deep Learning (= Neural Network)

input -> hidden layer -> output

  • Basic unit = neurons (nodes)
    • ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด brain์„ ๋ชจ๋ฐฉํ•ด์„œ ๋งŒ๋“ค์–ด ์ง„ ๊ฒƒ!
  • ML์˜ ์ข…๋ฅ˜ ์ค‘ ํ•˜๋‚˜
  • ๋งŽ์€ Data๋ฅผ ํ•„์š”๋กœ ํ•จ
  • input์ด ์ด๋ฏธ์ง€๋‚˜ ํ…์ŠคํŠธ์ผ ๋•Œ ๋ฒ ์ŠคํŠธ์ž„

์ฆ‰, AI ์ข…๋ฅ˜์— ML, ML์˜ ์ข…๋ฅ˜์— DL์ด ์žˆ์Œ!!!!

๋”ฅ๋Ÿฌ๋‹์€ ์–ธ์ œ, ์™œ ์‚ฌ์šฉํ• ๊นŒ?

  1. ๋ฐ์ดํ„ฐ๊ฐ€ ๋งŽ์„ ๋•Œ
  2. Access to processing power
  3. domain knowledge๊ฐ€ ๋ถ€์กฑํ•  ๋•Œ
  4. ์ž๋™์œผ๋กœ feature ์ถ”์ถœํ•ด ์คŒ
  5. ๋ณต์žกํ•œ ๋ฌธ์ œ์ผ ๋•Œ
    • Computer Vision (CV)
    • Natural Language Processing (NLP)

 

Percentron
= single-layer neural network

Neocognitron
= Multi-layer neural network (hierarchical multilayered architecture)

  • CV์—์„œ ์“ฐ์ž„

 

CV

CNN(Convolution Neural network)

Apply filters to generate feature maps

  • ํ•œ ํ”ฝ์…€๋กœ๋Š” ์–ด๋– ํ•œ ์ •๋ณด๋„ ์„ค๋ช…ํ•  ์ˆ˜ ์—†์Œ
  • Window(small area) ์‚ฌ์šฉ : context of image๋ฅผ ๋‹ด๊ณ  ์žˆ์Œ
  • input์—์„œ 4*4 patch์—๋‹ค๊ฐ€ ํ•„ํ„ฐ๋ฅผ ์ ์šฉํ•จ
    • 4*4 Filter => 16๊ฐœ์˜ different Weights๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Œ
  • ๋‹ค์Œ patch๋ฅผ ์œ„ํ•ด 2ํ”ฝ์…€ ์˜†์œผ๋กœ ์ด๋™ํ•œ๋‹ค
  • ์ด๋Ÿฌํ•œ patchy operation = Convolution

Applications

  • Facial recognition
  • self-driving vehicles
  • Automatic detection of tumors in CT scans
  • Deep fake

 

NLP (Natural Language Processing)

The ability for computers to understand the meaning of human language

  • Bag of words
  • Unstructured Data --> Structured Format

ํ•˜์ง€๋งŒ bag of words๋Š” ๋™์˜์–ด(synonyms)๋ฅผ ๊ตฌ๋ณ„ํ•˜์ง€ ๋ชปํ•จ
blue, sky-blue, aqua, cerulean๊ณผ ๊ฐ™์€ ๋‹จ์–ด๋ฅผ ์„œ๋กœ ๋‹ค๋ฅธ ๋‹จ์–ด๋กœ ์ธ์‹
์ด๊ฒƒ์„ ํ•˜๋‚˜์˜ single feature๋กœ groupํ™” ํ•˜๊ณ  ์‹ถ์Œ!

Word Embeddings (Word2vec)

  • ๋น„์Šทํ•œ ๋‹จ์–ด๋ฅผ ๊ทธ๋ฃนํ™”ํ•˜๋„๋ก feature๋ฅผ ๋งŒ๋“ ๋‹ค! (feature = mathematical meaning)
    ex. (King-man) + (woman - ?) -> Queen

Applications

  • Language translation (NMT)
  • Chatbots
  • Personal assistants
  • Review Analysis : Sentiment Analysis(๊ฐ์ •๋ถ„์„), Topic Model (LDA)

 

 

๊ทธ ๋ฐ–์—....

 

TTS (Text to Speech) : ํƒ€์ž…์บ์ŠคํŠธ
STT (Speech to Text) : ํด๋กœ๋ฐ” ๋…ธํŠธ, ๊ตฌ๊ธ€ ๋ณด์ด์Šค
TTI (Text to Image) : GauGAN bu Nvidia
TTV (Text to Video) : Wayne Hills

 

 


Explainability - Blackbox issue

  • Black box
    • Deep learning
      • Better for "What"
      • Highly Accurate predictions
  • Explainable AI
    • Tranditional machine learning
      • Better for "Why"
      • Understandable for humans

GIGO (Garbage in garbage out)

Gender&Race Bias