An introduction to AI (…and why you might avoid that term)

Previous attendees have said…

• 18 previous attendees have left feedback
• 100% would recommend this session to a colleague
• 100% said that this session was pitched correctly

Three random comments from previous attendees

• This was a great session. Thanks
• the essentials were there, whatever the care and social health focus; some of the information tidied up a few thoughts I had about AI
• Important for colleagues to have access to this sort of content.

Session materials

• all materials
• slides html / pdf

Welcome

• this session is 🌶: for beginners
• it aims to do two things:

1. to suggest that the term AI is troublesome
2. to introduce some of the different technologies that get lumped together as AI

A philosophical question

• do submarines swim?

Hype

• There’s a lot of hype about AI at the moment (see this graph)
• Underneath the hype, there’s a lot of genuinely exciting stuff going on too
• That exciting stuff is likely to have some impact on health and care work

Motive

1. The intelligence part of AI is as misleading as a swimming submarine
2. There are lots of different technologies that currently fall under the AI umbrella
3. Points 1. and 2. matter in a practical way because of the hype
   

What does AI mean to you?

HAL 9000

The Terminator

Generative AI image of the Pope’s coat

Siri and other personal assistants

Is AI…

• Over-hyped?
• Somewhere in between?
• Neglected?
• Other / don’t know

About this talk

Two linked problems:

• a worry about intelligence: based on the swimming submarine
• a worry about diversity: AI is several things, not just one thing

The Chinese room

Searle (1980)

“Suppose that I’m locked in a room and given a large batch of Chinese writing. Suppose furthermore (as is indeed the case) that I know no Chinese, either written or spoken, and that I’m not even confident that I could recognize Chinese writing”

However, he is supplied with a set of intelligible rules for manipulating these Chinese symbols

“火” is the opposite of “水”

“六” is more than “四”

Question

Does this poor bloke locked in a room understand the Chinese symbols?

Now suppose that we start asking him questions (in English):

Is “六” more than “四”?

If so, respond with “是”. Otherwise respond “不”

Question

• Is understanding the same thing as being able to produce output in response to input?
• Searle (1980) - this is the difference between strong and weak AI

Back to nice safe words

• we usually don’t worry too much about what words like intelligence, understanding, etc really mean
• for most purposes, understanding something, and doing that thing, pretty well overlap
• AI, unfortunately, is an exception
• big difference between producing output and understanding here

Why does this matter?

• Because the current conversation around AI does violence to our usual understanding of basic terms (like intelligence)
  • We need to do a bit of re-interpreting…
  • …particularly because AI can do the input-output part really well
• (side effect) The Chinese Room is an excellent way of understanding what’s going on inside some of the current tech

The tech

• AI = big umbrella term
• More specific terms:
  • Algorithms = rule-based ways of producing sensible output
  • Expert systems = more sophisticated expertise-based production of output
  • Machine learning = umbrella term for non-expertise-based production of output
  • Large Language Models = a massively-succesful sub-species of machine learning

So what’s an algorithm?

(Packard 1979)

• Algorithm = rule (roughly)
  • if something happens, do something
• made from expert input and evidence

An example algorithm

Related expertise-based tools

“See also…” references in indexes, library catalogues, wikipedia

. . .

• Brilliant 1996 Master’s dissertation looking at the state of “see also…” referencing in Ohio’s public libraries

How about something more complicated?

Part of the MYCIN inference system

• one problem with algorithms: how to handle conflicting information?
• An expert system - MYCIN (Shortliffe and Buchanan 1975)
  • designed to identify bacterial infections and suitable Rx
  • 600 rules, supplied by experts
  • asks users a series of clinical questions
  • combines the answers using a (fairly simple) inference system
  • able to manage some conflicting information - unlike simpler algorithms

Machine learning

• A next step: can we provide learning rules to a system, and let it figure out the details for itself?

https://commons.wikimedia.org/wiki/File:Supervised_machine_learning_in_a_nutshell.svg

This is supervised learning

• supervision = labelled observations used for training and testing
• Lots of health examples with promising results:
  • diabetic retinopathy (Mookiah et al. 2013)
  • ECG (Aziz, Ahmed, and Alouini 2021)
  • fractures, melanoma, …

A dataset downside

Fashion-MNIST dataset

Producing labelled datasets is hard:

• generally must be very large
• generally requires expert classification
• must be done with great accuracy
  • scale bar problem (Winkler et al. 2021)
• so dataset labelling is wildly expensive and thankless
  • Is there a way of doing something similar without spending trillions classifying everything in the world by hand?

Unsupervised learning

English-languge autocomplete suggestions

Unsupervised learning

Autocomplete sources

Unsupervised learning

German-languge autocomplete suggestions

Unsupervised learning

• No-one is writing a list of possible searches starting with “Large…”
• Nor are they classifying searches into likely/unlikely, then training a model
• Instead, the model is looking at data (searches, language, location, trends) and calculating probabilities
  • 2011 blog post
  • 2020 PR piece
  • 2020 build your own in JS
• The terminology gets confusing again at this point:
  • some describe this as deep learning
  • better to call this a language model

Large language models

What if we were more ambitious with the scope of our language model?

Transformer model architecture

• Find masses of language data
  • chatGPT uses basically the whole web before September 2021
• Build a model capable of finding patterns in that data
  • Attention model used in chatGPT (Vaswani et al. 2017)
• Allow the model to calculate probabilities based on those patterns
  • lots of work going on at present allowing models to improve in response to feedback etc

Large language models

• superb at generating appropriate text, code, images, music…
• but production vs understanding
  • e.g. hallucinations, phantom functions…
• training is extremely computationally expensive
  • questions about inequality and regulatory moating
    • no-one but FAANG-sized companies can afford to do this
  • training is also surprisingly manual

Ethics

• your web content, my model, my paycheque
• where’s the consent here?
• big serious worries about bias in some kinds of output
• rights violations via AI
• no settled questions around responsibility
• UK GDPR etc assume data is identifiable. That’s not true in LLMs.

Punchline

• On balance, while there’s hype here, there’s also lots of substance and interest
• LLMs have become much better at producing plausible output, across a greatly expanded area
• A strength: fantastic ways for those with expertise to work faster
• A danger: LLMs are great at producing truth-like output. Good enough so that some will be tempted to use them to extend their apparent expertise…
• But big serious legal and ethical trouble ahead - we’re not good at dealing with distributed responsibility

Thumbs-up for specificity

• many of the touted benefits are technology-specific
  • e.g. if we want to understand why decisions are getting made in a particular way, an expert system is better than a LLM
• we should probably start asking “what do you mean by AI” whenever we’re trying to make decisions about it

Conclusion

1. The intelligence part of AI is as misleading as a swimming submarine
2. There are lots of different technologies that currently fall under the AI umbrella
3. points 1. and 2. really matter because of the hype

Why the hype matters

• hype leads to perverse incentives and malfescience: call any rubbish AI, and get paid for it
• that means that both what we mean by AI, and what tech gets included, is extra-important at present - there’s an industry out there that’s profiting from blurring the boundaries

References

Aziz, Saira, Sajid Ahmed, and Mohamed-Slim Alouini. 2021. “ECG-Based Machine-Learning Algorithms for Heartbeat Classification.” Scientific Reports 11 (1). https://doi.org/10.1038/s41598-021-97118-5.

Mookiah, Muthu Rama Krishnan, U. Rajendra Acharya, Chua Kuang Chua, Choo Min Lim, E. Y. K. Ng, and Augustinus Laude. 2013. “Computer-Aided Diagnosis of Diabetic Retinopathy: A Review.” Computers in Biology and Medicine 43 (12): 2136–55. https://doi.org/10.1016/j.compbiomed.2013.10.007.

Searle, John R. 1980. “Minds, Brains, and Programs.” Behavioral and Brain Sciences 3 (3): 417–24. https://doi.org/10.1017/s0140525x00005756.

Shortliffe, Edward H., and Bruce G. Buchanan. 1975. “A Model of Inexact Reasoning in Medicine.” Mathematical Biosciences 23 (3-4): 351–79. https://doi.org/10.1016/0025-5564(75)90047-4.

Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. “Attention Is All You Need.” https://doi.org/10.48550/ARXIV.1706.03762.

Winkler, Julia K., Katharina Sies, Christine Fink, Ferdinand Toberer, Alexander Enk, Mohamed S. Abassi, Tobias Fuchs, and Holger A. Haenssle. 2021. “Association Between Different Scale Bars in Dermoscopic Images and Diagnostic Performance of a Market-Approved Deep Learning Convolutional Neural Network for Melanoma Recognition.” European Journal of Cancer 145 (March): 146–54. https://doi.org/10.1016/j.ejca.2020.12.010.