This document discusses challenges in building operational AI systems and provides guidance on how to approach AI development. It notes that while AI will significantly impact many aspects of life, building reliable operational AI systems faces hurdles such as a lack of high-quality data, problems being underspecified, tools not being designed for real-world use, and proof-of-concept systems failing to scale. The document offers design guidelines for AI such as ensuring ecological validity, being data-driven, making all aspects immutable and version-controlled, considering scaling from the start, and avoiding strong dependencies. It also provides advice on getting started with AI through self-study, using existing solutions, implementing non-critical test cases, and growing knowledge and processes
4. Peltarion - the operational AI platform
Technological disruptions
Internet
Cost of
Distributing Information
AI
1950 1960 1970 1980 1990 2000 2010
Transistor
Cost of
Arithmetic
Cost of
Prediction
5. Peltarion - the operational AI platform
History of Artificial Intelligence
Use data to automatically
learn to make predictions
Learn to both represent data
and make predictions
1950 1960 1970 1980 1990 2000 2010
Artificial
Intelligence
Science and engineering
of building intelligent machines
Machine
Learning
Deep
Learning
6. Sundar Pichai / CEO Google
“AI is one of the most important
things humanity is working on.
It’s more profound than, I don’t
know, electricity or fire”
7. Author: Owen Williams, The Kasparov Agency
Licensed under the Creative Commons Attribution-Share Alike 3.0 Unported license
“I could feel – I could smell –
a new kind of intelligence across
the table.”
Garry Kasparov
12. Detect the data types in CSV files with errors?
... ... ... ... ...
2012-04-23T18:25:43.511Z 12 54.4 283 Jennings Lane, Montgomery Village, MD 20886 Tree
2013-05-22T11:15:23.331Z Dog 98012.2345 9341 553 45
2015-01-03T 95 23.1 8529 West Pin Oak St. Copperas Cove, TX 76522 Sidewalk
2014-12-01T09:42:55.513Z 14 23 822 West Orange Dr. Long Branch, NJ 07740 Person
... ... ... ... ...
19. Modelling possible for non-experts
Support vector machine solution Deep learning solution
Example: Retinal disease diagnosis system
VS
20. Training time can be long and demanding
But lookups can still be done really fast*
Ready for real-world applications
* < 200ms even for advanced models
21. Real world ready or not?
Every technological shift has two phases
27. POC & Research
New tech
Experimentation
Machine Learning
Product development
Reliability & performance
Maintain & refactor
Software engineering
Different cultures
28. High risk, high reward
Sometimes it is easy,
sometimes there is no
solution at all
Illustration by Tony Oliver.
Licensed under Creative Commons Attribution-ShareAlike 2.0 Generic
29. Not enough data or
lack of quality
Underspecified problem
Photo by: Austin Mills
Licensed under Creative Commons Attribution-ShareAlike 2.0 Generic
30. Existing tooling is not
made for solving
operational AI problems
Mostly tilted towards research,
not development and solutions
38. Take nothing for granted - until you’ve tried it
Experiment & learn
1
Ecological validity
2
Be prepared to fail
3
39. Be data and use-case driven
Clear problem statement
1
No data - no fun
2
Background illustrations: User by Justiconnic & Data by priyanka from the Noun Project
40. Immutability - it is just easier that way
for any distributed system
1
Foundation for versioning
2
Background illustration: Protected Edit by Mun May Tee from the Noun Project
41. Version everything
Background illustration: Connection by Jason D. Rowley from the Noun Project
Tuning of deep learning parameters is not self-explanatory: Document changes!
Model history
1
Model used in serving
2
Version of data used
3
42. Think of scaling from the beginning
Training can take a long time
1
Experimentation - parallel jobs
2
Engineering debt of trying to make a non-scalable system scalable can be enormous
43. Know your GDPR
Photo by: Dennis van der Heijden
Licensed under Creative Commons Attribution-ShareAlike 2.0 Generic
44. Costs matter… sometimes a lot
Massive scaling on expensive GPU:s may quickly put an end to serious AI effort
Add monitoring
1
Hardware requirements?
2
Early stopping
3
45. Avoid glue code and strong dependencies
Use libraries only if you must
1
Wrap vital libraries
2
Semantic versioning
4
Create your own formats
3
Background illustration: Glue by David Marioni from the Noun Project
47. AI - from A-Z
Learn & have fun
Devote time for self-studies
Use existing solutions
Build demos
01
Get ready
Data & use case inventory
Spread knowledge
Start collecting missing data
02
Implement
One non-critical case with
business value
Existing vs your own tech?
Guidelines
03
Grow & improve
Spread AI-knowledge in
organisation
Processes & workflows
Recruit
04
48. Do not reinvent the
wheel each time. Think
in terms of a
framework early on.
Illustration by Loco Steve.
Licensed under Creative Commons Attribution-ShareAlike 2.0 Generic
53. Help us with data* to
train an AI beer model
Photo by: Darien Graham-Smith
* visit us in our booth and drink beer
daniel@peltarion.com // bit.ly/jfokus-ai
Thank you!