machine learning lecture 1 - edxmitx+6.86x+1t2019+type@... · 2019. 8. 20. · machine learning is...
TRANSCRIPT
![Page 1: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/1.jpg)
Machine Learning Lecture 1
![Page 2: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/2.jpg)
Machine learning is everywhere‣ Search, content recommendation, image/scene analysis,
machine translation, dialogue systems, automated assistants, game playing, sciences (biology, chemistry, etc), …
Learning to act: ex #3‣ Learning to play games via reinforcement learning ‣ Game of Go (branching factor ~ 250, length ~ 150) ‣ Beating a top human player was thought to be at least a
decade away ‣ AlphaGo (a deep neural network
model from DeepMind) - European master: 5-0 (15’) - Lee Sedol 4-1 (16’)
‣ Built on the basis of - expert play (supervised learning) - self play (RL) - search (local strategic choices)
![Page 3: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/3.jpg)
Machine learning: what is it?‣ A brief definitionMachine learning as a discipline aims to design, understand and apply computer programs that learn from experience (i.e., data) for the purpose of modeling, prediction, or control
![Page 4: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/4.jpg)
Prediction problems‣ About future events
‣ Also collision avoidance, monitoring, medical risk, etc.
Time
Mar
ket v
alue
![Page 5: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/5.jpg)
Prediction problems‣ About properties we don’t yet know
Convolutional Neural NetworkY LeCun
MA Ranzato
Object Recognition [Krizhevsky, Sutskever, Hinton 2012]
(Krizhevsky et al., 12’)
what is the image about?
“ML is very cool” what is it in Spanish?
would I like this movie?
soluble in water?
![Page 6: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/6.jpg)
‣ It is easier to express tasks in terms of examples of what you want (rather than how to solve them)
‣ E.g., image classification (1K categories)
Image CategoryConvolutional Neural Network
Y LeCunMA Ranzato
Object Recognition [Krizhevsky, Sutskever, Hinton 2012]
(Krizhevsky et al., 12’)
Convolutional Neural NetworkY LeCun
MA Ranzato
Object Recognition [Krizhevsky, Sutskever, Hinton 2012]
(Krizhevsky et al., 12’)
…
mushroom
cherry
…
Example: supervised learning
![Page 7: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/7.jpg)
‣ It is easier to express tasks in terms of examples of what you want (rather than how to solve them)
‣ E.g., image classification (1K categories)
‣ Rather than specify the solution directly (hard), we automate the process of finding one based on examples
Image CategoryConvolutional Neural Network
Y LeCunMA Ranzato
Object Recognition [Krizhevsky, Sutskever, Hinton 2012]
(Krizhevsky et al., 12’)
Convolutional Neural NetworkY LeCun
MA Ranzato
Object Recognition [Krizhevsky, Sutskever, Hinton 2012]
(Krizhevsky et al., 12’)
…
mushroom
cherry
…
Example: supervised learning
![Page 8: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/8.jpg)
‣ It is easier to express tasks in terms of examples of what you want (rather than how to solve them)
‣ E.g., image classification (1K categories)
‣ Rather than specify the solution directly (hard), we automate the process of finding one based on examples
Image CategoryConvolutional Neural Network
Y LeCunMA Ranzato
Object Recognition [Krizhevsky, Sutskever, Hinton 2012]
(Krizhevsky et al., 12’)
Convolutional Neural NetworkY LeCun
MA Ranzato
Object Recognition [Krizhevsky, Sutskever, Hinton 2012]
(Krizhevsky et al., 12’)
…
mushroom
cherry
…
Example: supervised learning
h
✓ ◆; ✓ =
![Page 9: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/9.jpg)
‣ It is easier to express tasks in terms of examples of what you want (rather than how to solve them)
‣ No limit to what you can learn to predict…
‣ Already in production for some language pairs (Google)
Is it real? ¿Es real?
English Spanish
Will it continue? ¿Continuará?
For how long? ¿Por cuanto tiempo?
… …
Example: supervised learning
=h�
; ✓�
![Page 10: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/10.jpg)
A concrete example‣ Learning to predict preferences from just a little data…
![Page 11: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/11.jpg)
A concrete example‣ Learning to predict preferences from just a little data…
-1 +1+1-1 ?, ?, ….
![Page 12: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/12.jpg)
A concrete example‣ Learning to predict preferences from just a little data…
-1 +1+1-1
action?
comedy?
romance?
top lead?
Spielberg?
x(1) = [1 0 0 1 1 . . . 0]T (feature vector)
?, ?, ….
![Page 13: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/13.jpg)
Supervised learning‣ Learning to predict preferences from just a little data…
-1 +1+1-1
x(1) x(2) x(3) x(4) x(5), x(6), . . .
Training set Test set
?, ?, ….
![Page 14: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/14.jpg)
Supervised learning
- ++
-
-
-x1
x2
![Page 15: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/15.jpg)
Supervised learning: training set
- ++
-x1
x2
![Page 16: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/16.jpg)
Supervised learning: test set
- ++
-
?
?
??
?
?
?
??
?
?
x1
x2
![Page 17: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/17.jpg)
Supervised learning: training set
- ++
-x1
x2
![Page 18: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/18.jpg)
Supervised learning: classifier
- ++
-x1
x2
h : X ! {�1, 1}
![Page 19: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/19.jpg)
Supervised learning: classifier
- ++
-
h(x) = +1h(x) = �1
-
-x1
x2
![Page 20: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/20.jpg)
Supervised learning: classifier
- ++
-
h(x) = +1
h(x) = �1
-
-x1
x2
![Page 21: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/21.jpg)
Supervised learning: classifier
- ++
-
h(x) = +1
h(x) = �1
-
-x1
x2
?
?
??
?
?
?
??
?
?
![Page 22: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/22.jpg)
Supervised learning: classifier
- ++
-
h(x) = +1
-
-
h(x) = �1
x1
x2
![Page 23: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/23.jpg)
Supervised learning: classifier
- ++
-
h(x) = �1
-
-x1
x2
![Page 24: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/24.jpg)
Supervised learning: generalization
- ++
-
?
?
??
?
?
?
??
?
?
x1
x2
h(x) = �1
![Page 25: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/25.jpg)
Supervised learning + ‣ Multi-way classification (e.g., three-way classification)
‣ Regression
‣ Structured prediction
h
✓ ◆= politics
h
✓ ◆= $1,349,000
h : X ! {politics, sports, other}
h : X ! R
h : X ! {English sentences}h
✓ ◆A group of peopleshopping at anoutdoor market
=
![Page 26: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/26.jpg)
Types of machine learning‣ Supervised learning
- prediction based on examples of correct behavior ‣ Unsupervised learning
- no explicit target, only data, goal to model/discover ‣ Semi-supervised learning
- supplement limited annotations with unsupervised learning ‣ Active learning
- learn to query the examples actually needed for learning ‣ Transfer learning
- how to apply what you have learned from A to B ‣ Reinforcement learning
- learning to act, not just predict; goal to optimize the consequences of actions
‣ Etc.
![Page 27: Machine Learning Lecture 1 - edXMITx+6.86x+1T2019+type@... · 2019. 8. 20. · Machine learning is everywhere ‣ Search, content recommendation, image/scene analysis, machine translation,](https://reader035.vdocuments.us/reader035/viewer/2022071415/610fb05563dbbe48b3230298/html5/thumbnails/27.jpg)
Key things to understand‣ Posing supervised machine learning problems
‣ Supervised classification
‣ The role of training/test sets
‣ A classifier
‣ A set of classifiers
‣ Errors, generalization