1. Big Data, Deep Learning, and Safety
Presentation for ACM Facility Safety, Calgary, Canada
01Sep2016
nets/ ACM safety presentation.ogv
Permission is granted to copy, distribute and/or modify ONLY the non-third-party content of this document under either:
The GNU Free Documentation License ( with no Invariant Sections, Front-Cover Texts, or Back-Cover Texts.
Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported License.

2. Computational Intelligence (CI) Description
Serious hobby interest - neural networks since 1988
peer reviews [journals, conferences, special requests] - stability & some control
IJCNN conference organisation - technical co- chair, publicity publications
elected Director of INNS Board of Governors (3 year term)
focus on "MindCode" - [architecture, functions, processes, "operating systems",
Other hobby interests
fundamental theoretical physics, astronomy, geology, history
Chasing crazy ideas
Career - chemical & mining industries, government research
industry - chemical plant operations, market research, business development

3. Big Data
Patterns versus rare anomalies
German power grid
Sociology

4. Big Data Description & Challenges
size & complexity always [have been, are, will be] issues
extreme [load, complexity, dimensionality, timeframe] beyond expert human capabilities
humans may need help with concept [identification, development, team-building, reporting]
highly heterogeneous & private datasets in working environments
Dimensionality reduction
privacy
HUMAN LANGUAGE - written, spoken, video
BIG DIMENSIONALITY - far more variables than data (DNA, genomics, proteomics, neuroinformatics)
Explaining your [approach, results] to [engineers, IT, managers, public]
The hard sciences are easy, the soft sciences are impossible

5. Big Data example Patterns versus Rare anomalies
Data availability, interactive requests
Supervised versus unsupervised learning
Clustering & Patterns - many approaches
Anomalie detection

6. Big Data example German power grid
Power grid - described by some as the most complex human-built system (communications, web?), while the human brain is sometimes called the most complex system in the universe
Approximate (Adaptive) Dynamic Programming (ADP) - self-[adaptive AND optimizing]. Finally a unification of two control theory communities !!!

7. Big Data example Sociology
fuzzy sets basis, correlational analysis later
sdf
The Bell Curve
Ghostwriter - Confabulation & composition

8. Deep Learning
Driverless cars
Checkers, Chess, Go
Social Media

9. Deep Learning description
GPU's as spark for the application of, and frenzy about, Deep Learning
size & complexity always [have been, are, will be] issues
extreme [load, complexity, dimensionality, timeframe] beyond expert human capabilities
humans may need help with concept [identification, development, team-building, reporting]
highly heterogeneous & private datasets in working environments
Dimensionality reduction
privacy
HUMAN LANGUAGE - written, spoken, video
BIG DIMENSIONALITY - far more variables than data (DNA, genomics, proteomics, neuroinformatics)
Explaining your [approach, results] to [engineers, IT, managers, public]
The hard sciences are easy, the soft sciences are impossible

10. Deep Learning example Driverless cars
Image [segmentation, recognition, cognition] is perhaps the critical component enabling driverless cars
Deep Learning has revolutionized this
Auditory analysis isn't being talked about (except sonar for object proximity) , nor is vibration (feel)
Smell & taste aren't usually associated with driving... and who has LIDAR (Laser Imaging, Detection, And Ranging) on their car? (Two eyes gives that)

11. Deep Learning example Checkers, Chess, Go
Collaborative-Competitive agents are a key characteristic of [biology, human environments] - DON'T UNDERESTIMATE this!
Classical games - initially algorithmic approach
Chess
Deep Blue versus Kasparov - Artificial Intelligence (AI)
later Computational Intelligence (an early version not grandmaster, used Evolutionary Computation)
Element of "surprise" Go
Go is to chess, as chess is to "tic-tac-toe"
only CI (Deep Learning including reinforcement learning)

12. Deep Learning example Social Media
Facebook & Google - bought up the best researchers/ companies in Deep Learning (Google got best of best - ~800M$ for ??)
Anticipate leap in [new, quality, powerful] capabilities
Does the machine know more about you than you do? (as a comparison to others, ranking of [comprehension, specialties, influence, ultimate "correctness and insight"])

13. What does this mean for Safety?
Coporate verbal, s, & reports
NASA
Computation Intelligence itself - safety risks
Hippocampal prosthesis

14. Safety example Corporate verbal, emails & reports
Information is often [plentiful (Big Data), heterogenoeous, disperse], but accessing it and using it is problematic
Use of [verbal, image, video] data becoming possible
Data [cull, organize, analysis, report] according to [objectives, needs]
Examples - Next plausible sentence, Ghost writer

15. Safety example NASA's Apollo 1&13, Challenger
Apollo 1 astronauts killed during routine training, Frank Borman "It was a failure of imagination by all of us"
Apollo 13 moon landing rescue following fuel cell failure. Amazing rescue, but what caused it?
Challenger O-ring failure
Could modern CI tools help with this type of problem?
Contrast with Medical Diagnosis systems - in some special applications can beat almost all human experts, but still not the best (hybridize, develop people), input to doctors rather than [directing, deciding]. FDA & other approvals of advanced technologies

16. Safety example Challenges arising from CI itself!
Black boxes - less of an issue with simple fuzzy systems, but even they can be problematic in complex applications.
Generalization failures leading to excursions
Unlike classical systems, there may be no mathematical theorems and proofs of stability
Application to challenges that humans cannot master & understand
How does one analyse a system that is continuosly [learning, evolving] in an environment that is doing the same?

17. Safety example - Human behaviours Hippocampal prosthesis
What is real?
Laguerre-Volterra kernels
Can this ever work? - anti-engineering and anti- Murphy's law

18. Conclusions

19. Computational Intelligence CI questions, thoughts
Almost all really advanced computing is bio- inspired
[Logical, rational, scientific] reasoning is a great tool, but it's not enough, and it may not be the right starting point (AI versus CI), ultimately fails
Is Deep Learning being done properly? (ordered derivatives vs backprop for complex architectures)
Gradient-free solutions (Cubature Kalman filtering, PSO, etc) - more powerful for control.
Stunning creativity of evolutionary computation [intuition, imagination]?
Anti-Engineering, anti-Murphy's Law in biological systems
"Universal function approximators" can do almost everything, but they do almost nothing well
Multiple conflicting hypothesis
You can't understand a science without knowing its history

20. Computational Intelligence Safety [Random, scattered] thoughts
Systems approaches - how can CI help?
[Fear, bother] of reporting, and are you interested in what people have to say from their own perspective? (no data into CI - nothing out)
Higher-level [conformance to specs, identify missing [checks, concepts, trends]?
Anomaly detection of rare combinations of [conditions, events] - as with [security, intelligence]
Safety is people - [objectives, requirements, design, training, operations], Is there anything to be taken from the USA program for vets?

21. Extra slides & themes

22. Big Data example Earthquakes & Lightening strikes

23. Safety example Underground mine automation

24. Safety example Fighter pilot ejection

25. Big Data example Stability of Memristor circuits
Memristors - the "fourth missing electrical circuit element" (1971 concept, following resitsors, capacitors, inductors)
Hewlett Packard development of memristors ?2006? - now putting 80% of r&d into "the machine" for business systems. Changes how hardware & software used.
Flash memory killer, never shut off computers?
Neuromorphic applications - neural networks
Mathematical proofs [existence, uniqueness, stability] of huge memristor circuits
FRACTIONAL ORDER CALCULUS !!!

26. Big Data example Engineering & Safety Design
Use of CI for optimizing designs
Geeting "good" answers for NP hard problems
Can we check final designs against specifications
Can we analyse final designs and modifications against specifications, standards and general requirements (engineering & safety perspectives)?

27. Deep Learning Howell's questions
Is backpropagation being properly applied? (Paul Werbos -> Don Wunsch, credit assignment)
This is currently accelerating - will it plateau soon?
What about marrying ADP & Deep Learning? (Google's blend of DL & reinforcement learning for computer games)
Yann LeCun "RNN's not way to go - too much labor intensive work", 2016 "Looking at RNNs"
Evolutionary Computational contributions
gradient-free solutions - eg Cubature Kalman Filters Simon Haykin McMaster (?A? now at Ford Research in Canada)
waiting for the next "new, big ideas"?
Meta-level understanding, re-use in different apps?