The document discusses various topics related to artificial intelligence including machine learning applications and demos, a toy machine learning problem, the history of AI from the 1940s to today, bias in AI systems, ethics, the technological singularity, and career opportunities in AI. It provides references and links to external resources for further reading on each topic. Live demonstrations are mentioned on computer vision applications and neural artistic style transfer.

1. Avilay Parekh
Sep 2018

2. What is AI?
AI & ML Applications (Live Demos)
A Toy ML problem
1940s to Today
Bias (More Live Demos)
Ethics
Singularity
Career

4. From 0:19 to 1:25

5. 0:50 to 1:20
2:50 to 3:09

6. 1:14 to 2:40

7. A human-made system that is able to learn.
The Roomba does not have to be programmed for each new room it operates in.
The Tesla is able to drive itself on a variety of different roads, traffic conditions,
and weather conditions.
Formal Definition by John McCarthy in 1956
AI involves machines that can perform tasks that are characteristic of human
intelligence.

8. AI: Artificial Intelligence
ML: Machine Learning
DL: Deep Learning
Ref: https://www.deeplearning.org

13. Ref: https://cloud.google.com/vision/
Demo: https://cloud.google.com/vision/

14. Ref: https://cloud.google.com/vision/

15. Ref: https://youtu.be/VOC3huqHrss

16. Ref: https://medium.com/artists-and-machine-intelligence/neural-artistic-style-transfer-a-comprehensive-look-f54d8649c199

17. Ref: https://arxiv.org/pdf/1605.05396.pdf

18. Ref: https://arxiv.org/abs/1710.10196

19. Two imaginary celebrities dreamed up by a Neural Network
Ref: https://arxiv.org/abs/1710.10196

20. Ref: https://research.googleblog.com/2016/03/deep-learning-for-robots-learning-from.html

21. Ref: https://connect.unity.com/p/pancake-bot

23. Predict whether an individual will play or not; given the weather conditions outside.
We have historical data about what this individual has done in the past.
Our job is to predict what they will do in the future.

24. Outlook Temperature Humidity Wind Play
Sunny Hot High Weak No
Sunny Hot High Strong No
Overcast Hot High Weak Yes
Rain MIld High Weak Yes
Rain Cool Normal Weak Yes
Rain Cool Normal Strong No
Overcast Cool Normal Strong Yes
Sunny Mild High Weak No
Sunny Cool Normal Weak Yes
Rain Mild Normal Weak Yes
Sunny MIld Normal Strong Yes
Overcast Mild High Strong Yes
Overcast Hot Normal Weak Yes
Rain Mild High Strong No
Rain Cool High Strong ?

25. Outlook Temperature Humidity Wind Play
Sunny=1 Hot=1 High=1 Weak=1 No=0
Sunny=1 Hot=1 High=1 Strong=2 No=0
Overcast=2 Hot=1 High=1 Weak=1 Yes=1
Rain=3 MIld=2 High=1 Weak=1 Yes=1
Rain=3 Cool=3 Normal=2 Weak=1 Yes=1
Rain=3 Cool=3 Normal=2 Strong=2 No=0
Overcast=2 Cool=3 Normal=2 Strong=2 Yes=1
Sunny=1 Mild=2 High=1 Weak=1 No=0
Sunny=1 Cool=3 Normal=2 Weak=1 Yes=1
Rain=3 Mild=2 Normal=2 Weak=1 Yes=1
Sunny=1 MIld=2 Normal=2 Strong=2 Yes=1
Overcast=2 Mild=2 High=1 Strong=2 Yes=1
Overcast=2 Hot=1 Normal=2 Weak=1 Yes=1
Rain=3 Mild=2 High=1 Strong=2 No=0
Rain=3 Cool=3 High=1 Strong=2 ?

26. Outlook Temperature Humidity Wind Play
1 1 1 1 0
1 1 1 2 0
2 1 1 1 1
3 2 1 1 1
3 3 2 1 1
3 3 2 2 0
2 3 2 2 1
1 2 1 1 0
1 3 2 1 1
3 2 2 1 1
1 2 2 2 1
2 2 1 2 1
2 1 2 1 1
3 2 1 2 0
3 3 1 2 ?

27. x1 =
outlook
x2 =
temperature
x3 =
humidity
x4 =
wind
y = play
w1
w2
w3
w4

28. Outlook Temperature Humidity Wind y_hat Play
1.w1 1.w2 1.w3 1.w4 b Close to 0 0
1.w1 1.w2 1.w3 2.w4 b Close to 0 0
2.w1 1.w2 1.w3 1.w4 b Close to 1 1
3.w1 2.w2 1.w3 1.w4 b Close to 1 1
3.w1 3.w2 2.w3 1.w4 b Close to 1 1
3.w1 3.w2 2.w3 2.w4 b Close to 0 0
2.w1 3.w2 2.w3 2.w4 b Close to 1 1
1.w1 2.w2 1.w3 1.w4 b Close to 0 0
1.w1 3.w2 2.w3 1.w4 b Close to 1 1
3.w1 2.w2 2.w3 1.w4 b Close to 1 1
1.w1 2.w2 2.w3 2.w4 b Close to 1 1
2.w1 2.w2 1.w3 2.w4 b Close to 1 1
2.w1 1.w2 2.w3 1.w4 b Close to 1 1
3.w1 2.w2 1.w3 2.w4 b Close to 0 0
3.w1 3.w2 1.w3 2.w4 b ? ?

29. 1.w1 + 1.w2 + 1.w3 + 1.w4 + b = Close to 0
1.w1 + 1.w2 + 1.w3 + 2.w4 + b = Close to 0
2.w1 + 1.w2 + 1.w3 + 1.w4 + b = Close to 1
3.w1 + 2.w2 + 1.w3 + 1.w4 + b = Close to 1
3.w1 + 3.w2 + 2.w3 + 1.w4 + b = Close to 1
3.w1 + 3.w2 + 2.w3 + 2.w4 + b = Close to 0
2.w1 + 3.w2 + 2.w3 + 2.w4 + b = Close to 1
1.w1 + 2.w2 + 1.w3 + 1.w4 + b = Close to 0
1.w1 + 3.w2 + 2.w3 + 1.w4 + b = Close to 1
3.w1 + 2.w2 + 2.w3 + 1.w4 + b = Close to 1
1.w1 + 2.w2 + 2.w3 + 2.w4 + b = Close to 1
2.w1 + 2.w2 + 1.w3 + 2.w4 + b = Close to 1

30. Linear Equation
Non-Linear Activation

31. ML Model
D

32. Model Parameters
Learnt by the system

33. Labeled Data
Software
Parameters
Feed in labeled training data, and out come the model
parameters

34. Input unlabeled data and model parameters; and the
software will output the prediction
Parameters
Unlabeled data
Software
Predictions

35. A 256x256 picture

36. A 256x256 picture Red Channel Green Channel Blue Channel
= + +

37. A 256x256 picture Red Channel Green Channel Blue Channel
Just 256x256
numbers
= + +

38. A 256x256 picture Red Channel Green Channel Blue Channel
Just 256x256x3 =
196,608 numbers
= + +

40. Same conceptual model - Linear function followed by an activation function
But the parameters were carefully selected - by hand!

41. X1 X2 Y
1 1 1
1 0 0
0 1 0
0 0 0
x1 +1
+1
-1.5
x2
1
y

42. X1 X2 Y
1 1 0
1 0 1
0 1 1
0 0 0
x1 ?
?
?
x2
1
y
No linear function followed by an
activation function can solve XOR!

43. X1 X2 Y
1 1 0
1 0 1
0 1 1
0 0 0
h1
h2
1
y
x1
x2
1
hidden layer
The solution was to add a
hidden layer - but that
didn't come until much
later..

44. Carefully hand-crafting weights was not going
to scale beyond toy logic gates.
To solve real world problems the "AI" system
was going to have to learn the weights itself,
based solely on input data.
In 1986 Geoffrey Hinton came up with the
Back propagation algorithm to do this.

45. Ref: https://beamandrew.github.io/deeplearning/2017/02/23/deep_learning_101_part1.html

46. Hinton introduces the idea of deep belief nets
Amazon AWS launches
Hadoop 0.1.0 is released
● Deep Learning needs lots of training data
● Processing lots of data requires Big Data technologies
● Big Data technologies require lots of computing power

47. Big Data + Cloud Computing + Deep Learning = AWESOMENESS!!

48. Machines competing against Humans has been a long
standing goal of AI.
In 1997 IBM Deep Blue defeated Chess champion Gary
Kasparov by analyzing all possible moves and selecting the
best move.
This will not work in the ancient game of Go. The number of
possible moves is greater than the atoms in the Universe!
In 2015 researchers in Deep Mind created a program that
defeated 18 time world Go champion Lee Sidol using a
technique in DL called Reinforcement Learning.
Huge waves in the AI community.
Must watch Netflix documentary
(https://www.netflix.com/title/80190844).

50. Simplest DNN
Each layer has 1000s of
nodes
A bunch of hidden layers
(usually 3 to 5)
Fully Connected = Each
node in one layer is
connected to all the nodes
in the next layer

51. https://www.mathworks.com/discovery/convolutional-neural-network.html
Usually applied to Computer Vision type problems
Most mature in terms of research

52. Usually applied to sequential data, e.g., daily temperatures, language models, etc.
Very powerful and gaining rapid traction
Ref: http://colah.github.io/posts/2015-08-Understanding-LSTMs/

53. Usually applied to generate images
Recently has been applied to generate other artifacts like 3D drawings, fonts, etc.
Ref: https://medium.freecodecamp.org/an-intuitive-introduction-to-generative-adversarial-networks-gans-7a2264a81394

54. Usually applied to train an "agent" to act in a certain way
Very exciting field with a lot of new research being done daily
Ref: http://karpathy.github.io/2016/05/31/rl/

56. Language Modeling
Assign probabilities to sentences, i.e., predict which word is most likely to follow
some sequence of words in a sentence.
Used in Speech2Text e.g., by Siri, Alexa, Ok Google, etc.
At its core is the ability to represent words as numbers. But how?

57. Demos
https://cloud.google.com/speech-to-text/
https://cloud.google.com/translate/
I like Gujarati because I love my mom

58. In 2013 researchers led by Tomas Mikolov at Google trained their model on a
billion+ Google News articles in English.
Each word in the English language is represented by a vector of 300 numbers.
Since then this technique has been applied to multiple languages including
Hindi and Gujarati.
Word2Vec has been instrumental in recent advancements in Language
Modeling.
Original Paper:
Efficient Estimation of Word Representation in Vector Space (https://arxiv.org/abs/1301.3781)

59. The model did something amazing! It learnt relationships between word pairs like
country-capital, man-woman, verb-tense. ALL ON ITS OWN!!
Ref: https://www.tensorflow.org/tutorials/representation/word2vec

60. father : mother :: doctor : ?
man : woman :: computer programmer : ?

61. father : mother :: doctor : nurse
man : woman :: computer programmer : homemaker
It learnt to be sexist! Again, ALL ON ITS OWN!!
Original Paper:
Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings
(https://arxiv.org/abs/1607.06520)

62. Ref: https://www.bbc.com/news/technology-33347866

63. Very active field of research
In May 2018 MS announced that they are working on bias identification tools
In May 2018 Facebook announced Fairness Flow - their tool to identify bias
In Sept 2018 Google announced What-If to analyze ML performance
In Sept 2018 IBM launched Fairness 360 Kit
But all these tools can only identify known bias like race, gender, age, etc. What
about our latent biases? E.g., in 1950s gender bias was the norm.

65. Ref: https://www.theregister.co.uk/2017/06/20/tesla_death_crash_accident_report_ntsb/
In AV mode crashed into a white trailer truck mistaking it for super bright sky killing
the car driver

66. Ref: https://internetofbusiness.com/uber-safety-driver-watching-video-before-fatal-crash-say-police/
Killed a pedestrian
who was jaywalking
across the road at
night
Driver was watching
TV on his phone

67. Ref: https://www.bloomberg.com/news/articles/2018-03-31/tesla-says-driver-s-hands-weren-t-on-wheel-at-time-of-accident
Crashed in a highway barrier
killing its driver
Driver ignored the car's
warning for 6 seconds before
the car finally crashed

68. A

69. B

70. A Do Nothing
B Pull the switch

71. A
B

72. A Do Nothing
B Pull the switch

73. A
B

74. A Do Nothing
B Pull the switch

75. No right answer
Depends on the individual making the decision and the context of the decision
Well known problem in studied by students in ethics since 1967
Wikipedia link for details and variants -
https://en.wikipedia.org/wiki/Trolley_problem
In Autonomous Vehicle settings we need to grapple with these kinds of issues

76. MIT Media Labs Moral Machine asks humans to answer similar questions
http://moralmachine.mit.edu/
National Highway Traffic Safety Administration issued AV safety guidelines
https://www.nhtsa.gov/technology-innovation/automated-vehicles-safety
Car companies rely on simulations to get certified
CARLA (http://carla.org/) open source self-driving simulation software
Apollo (http://apollo.auto/) open source self-driving hardware sensors
AirSim (https://github.com/Microsoft/AirSim) open source self-driving and self-flying simulation software

79. No, serious scientists have been talking about this phenomenon since 1958
including Von Neumann and Stephen Hawking
Technological progress has been advancing in leaps and bounds; but limited by
human intelligence which has pretty much remained the same for a millennia
Lets say we build an AI that is slightly more intelligent (capable) than humans
It will then figure out a way to build other AIs that are more capable than itself, who
in turn will build next gen AIs who are even smarter…
Eventually AIs will far surpass any and all human cognitive abilities giving them
god-like power over us

80. 1. A robot may not injure a human being, or, through inaction, allow a human
being to come to harm
2. A robot must obey the orders given by human beings except where such orders
would conflict with the First Law
3. A robot must protect its own existence as long as such protection does not
conflict with the First or Second Laws

81. 0. A robot may not harm humanity, or, by inaction, allow humanity to come to harm
1. A robot may not injure a human being, or, through inaction, allow a human
being to come to harm
2. A robot must obey the orders given by human beings except where such orders
would conflict with the First Law
3. A robot must protect its own existence as long as such protection does not
conflict with the First or Second Laws

82. Academic Think Tanks like Future of Life Institute are researching this topic
Industry is mostly ignoring the prospect of Singularity
Very big and completely unpredictable technology breakthroughs are needed
before such a Seed AI can be created
A very interesting discussion on Feb 2018 moderated by Neil deGrasse Tyson
https://www.youtube.com/watch?v=gb4SshJ5WOY

84. Computer Science
Linear Algebra
Calculus
Statistics & Probability

85. Hardware
Integrated Chip Design
Sensors
Consumer Products
Software
Low (Operating System) level APIs
ML libraries
ML Software As A Service
Applications

88. Data:
https://drive.google.com/open?id=1RP27WbpctWKL5vsRISb7FQma6oKgcQwh
Minimize J => Find the values of weights for which J is minimum

89. Entire dataset of (say) 1000 samples
Calculate the gradient over the entire dataset
Update all the weights

90. .
.
.
.
Do this for multiple iterations until the weights converge

91. x1 =
outlook
x2 =
temperature
x3 =
humidity
x4 =
wind
y = play
w1
w2
w3
w4

92. x1 =
outlook
x2 =
temperature
x3 =
humidity
x4 =
wind
y = play
a1
a2
a3
hidden layer
v
u

93. x1 =
outlook
x2 =
temperature
x3 =
humidity
x4 =
wind
y = play
a1
a2
a3
hidden layer
v
u

95. w
An easy way to calculate the gradients

96. a δ
w
w
No need to derive the entire cost function at once
Just consider the two nodes in question
a = forward propagation
δ = back propagation

97. a2
a1
a3
a
w1
w2
w3
δ
δ1
δ2
δ3
u1
u2
u3

98. δ
δ1
δ2
δ3
u1
u2
u3

99. δ
δ1
δ2
δ3
u1
u2
u3
δ4
δ5
δ6
v
As we go further back into the network, more and more weights are multiplied with
each other

100. δ
δ1
δ2
δ3
u1
u2
u3
δ4
δ5
δ6
v
δ8
δ9
δ7
w

102. As we go further back into the network, weights are multiplied with each other
If the weights are small (between 0 and 1), they become even smaller → Network
does not learn anything
If the weights are big, they become even bigger → Network learns nonsensical
things

103. .
.
.
.
What if our dataset has millions of samples?
Each small weight update has to wait until all the samples are processed
This can take a loooong time to converge!

104. Divide the entire dataset into mini-batches with T samples in each mini-batch
Calculate the gradient over each mini-batch
{1} {2} {t}...
Update the weights after each mini-batch

105. Do this for multiple epochs until the weights converge
…
…