1. AI, ML.Net,
GPT, Stable-
Diffusion
A leap around AI with
Dennis Vroegop

2. Don’t be afraid AI will steal your job…
Be afraid the other guy/girl who knows about
AI will steal it

3. A bit of history
Just adding some context

4. 1956
At Dartmouth Conference, John McCarthy claims the term “AI” AI is born!

5. 1980-1990
Expansion due to good hardware,
enabled neural networks

6. 1990-2000
The rise of the internet
enabled huge scale up of
resources.

7. 2000-2010
Deep learning and machine learning makes AI much more powerful

8. 2018 - Today
OpenAI hugely improves communications, AI for the masses.

9. But… why?
What are we going to do with it?

10. 2 billion cameras
connected to AI
• (NVIDIA said that by 2025 there will
be 2 billion video cameras in cities
around the world, they want to
connect those to gather data about
flow of traffic, crime and so on)

11. 25% of the companies will
dedicate workers to oversee AI
systems
• (Gartner said, “AI continues to drive
change in how businesses and
governments interact with
customers and constituents”. AI is
already heavily used in spotting
breast cancer, in such a way that this
can be detected a year earlier than it
would have been normally
diagnosed)

12. AI Market will be over
$60 billion in 2025
• (Research company Tractica says the
AI market will skyrocket from $1.4
billion in 2016 to nearly $60 billion
in 2025, but markets related to AI
will grow as well. For instance,
Machine Learning as a Service will
be worth $19.9 billion by the same
year)

13. Smart Voice Assistants will be
in 15 billion devices in 2025.
• (RBC Capital Markets predicts Alexa
alone could be worth $20 billion
2025, both from sales and spending
on the Amazon AI platform)

14. AI in driverless cars could save 300,000
lives in America each decade
• (NVIDIA, Alphabet’s Waymo and
Tesla are already using computer
vision and deep learning
technologies to allow vehicles to
both react to what they see and learn
from their experiences. We’ll see a
huge decline in number of traffic
fatalities on the road)

15. 30% of companies could have
their revenues reduced
• (Gartner predicts that companies
that perform software services right
now could become irrelevant, or a
least less useful, once machines start
figuring out how to do things better
and faster than they can. AI will
cannibalize revenues from some
current market leaders)

16. 38% of jobs in the U.S. could be
vulnerable to AI by the early 2030s.
• (Price Waterhouse Coopers says that
the most at-risk jobs are in
transportation and storage,
manufacturing and wholesale and
retail)

17. Building a neural
network.
By hand…

18. Meet
Morris…
Our 9 year old Labradoodle,
who is afraid of water.

19. Rain in mm
and outside
time in
minutes

20. Other factors
• Have I already had my
steps today?
• What time of day is it?

21. Quantization
Rain (in mm) < 5 5-10 > 10
1 2 3
Steps done < 5000 5000-10000 >10000
1 2 3
Time of day Morning Afternoon Night
1 3 2

22. Input A
Input B
Input C
Output

23. A
B
C
Input A
Input B
Input C
Output

24. A
B
C
Input A
Input B
Input C
Output

25. A
1
B
2
C
3
Input A
Input B
Input C
Output
1
2
3
4
3
5
7
6
8
9
1
2

26. 6
30
72
1
1
1
282
1
2
3
4
3
5
7
6
8
9
1
2

27. 18
60
144
3
2
2
570
1
2
3
4
3
5
7
6
8
9
1
2

28. End results..
Walking time 5 – 10 minutes 10 – 20 minutes > 20 minutes
< -1.5 -1.5 – 0.8 > 0.8

29. 2,5
1
-0,5
2
1
3
-0,5625
-1
-1
0,5
0,5
-1
0,5
-0,75
0,5
-0.75
-0.75
2,5
1

30. Party time!
We solved it!

31. How does it work?
I mean, internally…

32. “Old school”
“Knowledge systems…”

33. Introducing the neuron
The hero of our story

34. A
B
C
𝑆 =
𝑖=0
𝑛
𝑤

35. But a whole lot
of them…
• This simple model has:
• 8 inputs
• 3 hidden layers, each having 9
neuros
• 4 outputs
• 270 lines, thus weights
• 27 bias values
• In total 297 parameters…

36. What???
• Input: the number from your
model
• Bias: just a number to add to it
(weight)
• Activation Function: a simple
math function to apply
• Next Layer: do it all over
again, or use it as output
𝑆 =
𝑖=0
𝑛
𝑤

37. Activation functions
• Binary step: 0, if X < 0, or 1 if X > 0
• Linear Activation: f(x) = x
• Non-linear Activation: f(x) = sin(x)
• ReLU (Rectified Lineair Unit): f(x) = max(0,x)
• Parametric (ReLU): f(x) = max(ax, x)
• And so on…

38. Workflow…
(part 1)
• Create a model
• Set up number of input neurons
• Define number of layers
• Define number of neurons per layer
• Define number of output neurons
• Define initial weight / activation functions
• Define optimization model (more on this
later)

39. Workflow…
(part 2)
• Feed data into system
• Run through all layers
• Check outcome with known-results
• Tweak weights in neurons
• Do it all over again

40. Optimization
models
• There are hundreds of models
• Pick one that is ‘good enough’
• Each scenario requires its own

42. Machine
learning: it’s a
process, not a
tool!

43. WORKFLOW
Ask
Data
Clean
Train
Evaluate
Retrain

44. ASK THE
RIGHT
QUESTION
• Rephrase your question to
match ML paradigm
• Make it obvious
• Think of the genie in the lamp

45. FIND THE
RIGHT DATA
• Big data is the most-used
source
• Use sample data or pre-
selected data
• OpenData initiative
• In NLD: CBS data
• Google for ML Datasets and
be amazed

46. CLEAN UP
THE DATA
• Clean up missing values
• Substitute
• Remove rows
• Calculate medians / averages / and so on
• Remove outrages values
• Clip data

47. FIND THE
RIGHT
MODEL
• Anomaly detection
• Regression
• Clustering
• Two-class classification
• Multi-class classification

48. Let’s build
something
… and make it good!

50. Cats and dogs...
"Hello world" for AI Image Classification

51. Demo time!
Show me the code

52. ChatGPT
… or: the next step?

53. ChatGPT
• Chat… it’s about, uhm, chatting?
• Generative: it generates something
new
• Pre-trained: it has been trained
before
• Transformative: the model behind it
all

54. Generative
• The system generates
something new.
• Other options are:
• Predictive
• Classification
• … well, we’ve talked about
those

55. Pre-Trained
• It has been trained on a large
dataset
• Training was unsupervised
• Then finetuned on known samples
• Other options:
• Supervised learning: labeled
data
• Unsupervised. Used for
clustering
• And others…

56. Transformative
• The model.
• Lots of jargon about dual layer, each
with 2 sub-layers, that feed on each
other…
• If you really, really, want to
know:
[1706.03762] Attention Is All You
Need (arxiv.org)

57. Some numbers
• Trained on 45 terabytes of text
• 1.5 billion parameters
• Since released, answered over 1
billion questions
• Backed by Microsoft for over
$10.000.000.000
• A single training-run costs
$12.000.000
• Hit 1 million users in the first 5
days….

58. A demo
conversation

59. Stable Diffusion
Reverse AI, making new art

60. How it works:
• Generate non-uniform random noise
• Feed the keywords in a reverse
neural network
• Work back to see if parts of the
image fire certain regions
• Use an AI system to de-noise the
image
• Repeat de-noising until satisfied.

61. First we
generate
random
noise.

62. Then we
look for
boundaries.
(that satisfy our neurons)

63. The image
becomes
clear.

64. Almost
there…

65. Final result!

66. (Yes, I have
oversimplified
it massively…)

67. Try it out!
• Mage.Space. https://mage.space
• MidJourney Discord,
https://www.midjourney.com
• Stable-Diffusion running locally.
https://github.com/AUTOMATIC
1111/stable-diffusion-webui
• Dall-E (works slightly different,
same results)
https://labs.openai.com/

68. Conclusion
And thanks for listening!
Reach me at
dennis@vroegop.org