Generative Artificial Intelligence: How generative AI works.pdf
Machine Learning and AI, by Helena Deus, PhD
1. Machine learning
and AI
Dr. Helena F. Deus
Women in Tech Summit
Philadelphia, April 2018
Photo by François-Dominique / CC BY-SA 4.0
2. | 2Elsevier Labs
Machine learning is a field of computer science that
gives computer systems the ability to "learn" with data,
without being explicitly programmed.
Deep
Learning
Machine
Learning
AI
3. | 3Elsevier Labs
About Elsevier
• 130 year old company, HQ in Amsterdam
• 2500 scientific journals (e.g. Cell, Lancet) and 30 000 e-
books (e.g. Gray’s Anatomy)
• Today, a global information analytics business with a
mission to 1) advance healthcare; 2) enable open
science and 3) improve professional performance
Only great
science shall
pass
4. | 4Elsevier Labs
Gender distribution at Elsevier
35%
68%
54%
63%
31%
45%
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
Technology Non Technology All
Gender distribution at Elsevier
Female Male NA
Elsevier's FTE
gender
distribution is:
Female: 54%
Male: 45%
Tech industry average is 25%
Elsevier has a
unique market
position with over
10% more women
in tech roles than
industry average.
This can be used
for recruitment
purposes.
For open positions: Patrick Irwin (p.irwin.1@elsevier.com), https://www.elsevier.com/about/careers/technology-careers
5. | 5Elsevier Labs
About me
• Data Scientist
• BS in Biology, PhD in Bioinformatics
• Deep learning user for a little over a year
• Passionate about using AI for solving health care
WHAT MY FRIENDS THINK I
DO
WHAT I REALLY DO
8. | 8
A brief history of machine learning and AI
1840:Comput
ers can be
programmed
(Ada
Lovelace)
1950: Turing
test (Alan
Turing)
1952:
English-like
programming
languages
(Grace
Hopper)
1956:
"Artificial
intelligence"
is coined
(John
McCarthy)
1957: First
artificial
neural
network
(Frank
Rosenblatt)
1958:
Logistic
regression
(David Cox)
1969: Apollo
11 - learn low
and high
priority tasks
(Margaret
Hamilton)
1970: “AI
winter”
caused by
inflated hype
9. | 9
1982:
Recurrent
Neural Nets
(John Hopfield)
1993: Modern
Support Vector
Machines
(Corinna
Cortes)
1999:
Convolutional
Neural Nets
(Yann LeCun)
2006:
ImageNet (Fei-
Fei Li)
2011: IBW
Watson beat
humans in
Jeopardy
2012: Coursera
AI course
(Daphne Kohler,
Andrew Ng)
2014:
Facebook
publishes
DeepFace
2016: Google's
AlphaGo beats
humans in Go
A brief history of machine learning and AI
10. | 10Elsevier Labs
Big “Structured” Data
2 billion: number of
facebook users
82 million: amazon
reviews
14 million: labelled
ImageNet
12. | 12
How gradient descent works
NEEDS
IMPROVEMENT
ACCEPTABLE
IDEAL
KEEP TRYING
Cost or Loss Function: How far from
reality is the prediction
13. | 13
Regression(s)
If they visited 200 times, how much cash
would they spend?
Regression: pick the line that minimizes the
distance between the points and the line
http://scikit-learn.org/
http://colab.research.google.com/
14. | 14
Classification with Support Vector Machines
New flower has [6.2, 2.9, 4.3, 1,3] – can you
tell me the species?
15. | 15Elsevier Labs
Neural networks are easy with linear algebra
A
B
C
D
E
A
B
C
D
E
A
B
C
D
E
Back Propagation!
distance from target is 0.6
0.2
0.2
0.2
https://keras.io/
17. | 17Elsevier Labs
Deep Learning - neural networks with a lot of layers
https://www.cs.toronto.edu/~frossard/post/vgg16/
Convolutional Neural Networks (CNN) Generative Adversarial Networks (GAN)
“car”
https://towardsdatascience.com/gan-introduction-and-implementation-part1-implement-a-
simple-gan-in-tf-for-mnist-handwritten-de00a759ae5c
For you to Google: MNIST CNN Keras
Good website: https://machinelearningmastery.com/
22. | 22Elsevier Labs
Word Embeddings and Neural Networks
I
am
having
a
lovely
time
here
in
Philadelphia
positive
negative
0.1
0.9
Word2Vec
https://erikbern.com/2015/09/24/nearest-neighbor-
methods-vector-models-part-1.html
23. | 23Elsevier Labs
All mice were maintained in a temperature controlled (22 ± 2 °C) environment
12-h light 12-h dark photocycle and fed rodent chow meal .
The mice were individually placed into an acrylic cylinder (25 cm height 10 cm
diameter) containing 8 cm of water maintained at 22–24 °C
Cold mice and Cancer Research Deus et al 2017, IEEE
Training set: 480 sentences ; Train/Test split: 70/30; <1 min training time
Matching phrases (eg mice .. kept):
24.6% False
Discovery Rate
Using Neural Networks:
4% False
Discovery Rate
25. | 25Elsevier Labs
Why you should be
concerned about AI
“Ill-conceived mathematical
models
now micromanage the economy,
from advertising to prisons.”