Talk for MK Geek Night, 23 Sep 2021
AI means more hype, more technology, more future - and more money! But what actually is it? In this talk, Doug will explain what people mean by artificial intelligence and machine learning, what sort of problems they can solve, and how they do it. We'll see a range of examples where they're being used, and look at how it goes well and how it goes wrong, from entertaining AI weirdness to serious algorithmic bias. You won't end up being able to implement techniques like Support Vector Machines or Generative Adversarial Networks (unless you already could) but you should end up with a better idea of what the people who can are up to.
2. Who I am
Photo (CC)-BY-SA Bill Bertram https://commons.wikimedia.org/wiki/File:ZXSpectrum48k.jpg
3. Quick quiz:
How much do you
know about AI?
A. Nothing special
B. I’m interested
C. I’ve applied it myself
D. I have strong views
about the use of
gradient descent for
hyperparameter
optimisation in training
deep convolutional
neutral networks, or at
least, I fully understood
this sentence.
Photo by Maximalfocus on Unsplash
4. 1. What is AI?
2. How does it work?
3. How does it go wrong?
a) In entertaining ways
b) In serious ways
4. How does it go well?
5. 1. What is AI?
2. How does it work?
3. How does it go wrong?
a) In entertaining ways
b) In serious ways
4.V How does it go well?
6. Can machines
think?
• Turing Test:
Can you make a machine
humans can’t tell is a
machine?
• Eliza (algorithm)
• Chatbots (machine learning)
• GPT-3 (neural network)
Photo (CC) BY-SA Antoine Taveneaux https://en.wikipedia.org/wiki/Alan_Turing#/media/File:Turing-statue-Bletchley_14.jpg
9. Photo by Timelab Pro on Unsplash
• Games
• Chess, Go, video games
• Optimisation
• Distribution, manufacturing, retail, finance
• Complex control systems
• Robots, drones, self-driving cars
• Recommenders
• Amazon, YouTube, Netflix
• Speech recognition, natural language processing
• Siri, Alexa, Hey Google
10. 1. What is AI?
2. How does it work?
3. How does it go wrong?
a) In entertaining ways
b) In serious ways
4. How does it go well?
12. Photo by Anika Huizinga on Unsplash
More like this
(supervised learning)
What have we here?
(unsupervised learning)
Get a high score
(reinforcement learning)
18. By Petty Officer Photographer Jay Allen - https://www.royalnavy.mod.uk/news-and-latest-activity/news/2021/may/20/210520-carriers-at-sea-and-strike-warrior,
OGL 3, https://commons.wikimedia.org/w/index.php?curid=105562576
Fix bugs in sorting algorithms Delete the list to be sorted
(no items = nothing out of order)
Minimise difference between generated
generated and target output Delete the target output
Design an optimal lens It’s 20 m thick
Land aircraft on carrier So fast the braking cable force overflows
to zero
Bad optimisation
19. 1. What is AI?
2. How does it work?
3. How does it go wrong?
a)In entertaining ways
b)In serious ways
4. How does it go well?
22. BMJ 2020;369:m1328
7 April 2020
• 232 predictors
• Poorly reported
• High risk of bias
• Two worth validating
Nature Machine Intelligence
3, 199–217 (2021)
• 2,212 studies
• 415 after initial screening
• 62 after quality screening
• None of clinical use
Photo by Fusion Medical Animation on Unsplash
They used AI to help with Covid …
… it wasn’t good.
23. MQ-0 Reaper over Afghanistan
By Lt. Col. Leslie Pratt - commons file, Public Domain, https://commons.wikimedia.org/w/index.php?curid=68095681
New capabilities to do bad
Increasing inequality
Unemployment
25. Superintelligence
(CC)-BY William Clifford https://flic.kr/p/37PkGN
The AI does not hate you,
nor does it love you,
but you are made out of atoms
which it can use for something else.
26. 1. What is AI?
2. How does it work?
3. How does it go wrong?
a) In entertaining ways
b) In serious ways
4. How does it go well?
28. GANs
generator vs discriminator
Cats from http://thesecatsdonotexist.com/ https://devopstar.com/2019/02/25/generating-cats-with-stylegan-on-aws-sagemaker
30. Ethical AI
• Fairness
• Accountability
• Sustainability
• Safety
• Transparency
Leslie, D. (2019). Understanding artificial intelligence
ethics and safety: A guide for the responsible design
and implementation of AI systems in the public sector.
The Alan Turing Institute.
https://doi.org/10.5281/zenodo.3240529 Photo by Hendo Wang on Unsplash
43. How to use a neural network
1. Gather data
2. Train your
network
3. Use it on new data
4. Profit
Photo by Joshua Lanzarini on Unsplash
44. • It isn’t magic
• It depends on the data
• It has got way better
• It can go wrong
• in entertaining ways
• and in really bad ways.
• It will get better
So pleased to be here giving a talk.
Love giving talks, first in nearly 3y
This little guy here was put in by AI
Data scientist, digital transformation leader, researcher, teacherInterested in human & machine learning for decades
Turing Test Aka The Imitation Game, starring Benedict Cumberbatch
PowerPoint suggested these icons. Woah.
What coes AI do well now?
PowerPoint suggested these icons. Woah.
More like this – known data, cats/dogs: prediction, generation, pattern recognition, anomaly/noveltyWhat have we here – unknown data, find patterns, group this data, classify, categorise our customers
High score – a way of keeping score, explore/exploit: maximise revenue, minimise inventory, improve completion rate
Cells like in Excel
Numbers in Input are your data, photo pixels, Output, cat or dog
Adjust the numbers until the inputs give you the right output
Many more cells, more layers, arrows
I trained a NN on English placenames
It generated some fun ones
Janelle Shane
Inktober, creative prompt, draw every day for October
NN-generated prompts
Learn a locomotion strategy, fast. Makes a tall thing that falls over.
Jump as high as you can. No not rolling over. Highest point your lowest bit reaches. It falls over & pole vaults, not jumping.
YouTube, what sucks in your attention, has lots of vortexes that suck people in, tries them out on you. Ends up with the hard stuff on conspiracy theories, radicalization, etc
Facebook, socially meaningful engagement = most controversial posts, from friends & family, stuff you can’t help but engage on, and pull in reinforcements. I left.
Job applications, hard work, biased.
Train NN, prev applicants, who appointed.
Racist! Delete ethnicity. Names. Career paths, univs.
AI won’t stop racism, the people doing it have to work hard at stopping.
Sometimes it’s just not very good.
https://www.bmj.com/content/369/bmj.m1328https://www.nature.com/articles/s42256-021-00307-0
gig economy needs AI
How could really good AI be bad? Artificial general intelligence, not narrow
Self-improvement.
Paperclip maximiser- optimise manuf process, get smarter
Social intelligence, better deals with suppliers, sales
Far fetched, but so are the AIs below
Problem this week? No.
Worth looking at? Yes. And people are.
Chess used to mean intelligent! DeepBlue vs Kasparov 1990s
AlphaZero knew almost nothing, learned it from the game.
Chess, Shogi, Go
Huge compute effort to train, much less to play.
Generative adversarial network
Images, then cue generator in with prompt images
Deepfakes
Art!
Proteins DO everything, nanobots
Shape is everything
DNA, amino acids, protein
Lots of DNA seqs, fewer shapes
Transformatory
Fairness in data, design, outcome, implementation
Accountability before and after
Sustainability – economic, environmental, social (license to operate)
Safety – accuracy, reliability, security, robustnessTransparency – explainable AI, openness
There is, as the man used to say, one more thing. Example of a GAN applied to the video domain with both a motion prompt and a separate photo prompt
Combine with audio track to produce composite video
There is, as the man used to say, one more thing
If you can play chess, you’re intelligent
Mid40s, Turing theorized how computers could play chess. 1949, Claude Shannon at Bell Labs published paper w description of how do it. 1950, Turing made an actual algorithm. No suitable machine, so manually calculated, 30 min per move. Algorithm lost, history made, paper in 1953.
https://www.pcworld.com/article/2036854/a-brief-history-of-computer-chess.html#slide3
1958IBM programmer Alex Bernstein playing his chess program at the console of the 704 mainframe. Bernstein told the computer what move to make by flipping the switches on the front panel. The program took about eight minutes to calculate each move.