Dissertation - AI, Bias, Alignment & Other Topics

Slide 1

AI, Bias, Alignment



(& Other Topics!)


Slide 2

Let’s start with an old story… Image

1833 Alexander Pushkin visited Uralsk (https://adebiportal.kz/en/news/view/alexander-pushkin-215-anniversary__8687)
In the same year he published Eugene Onegin as a single volume


Slide 3

Image Towards AI: Andrey Markov (1856 - 1922)

Developed Markov Chains - key to many disciplines even today (Google “Monte Carlo Markov Chains (MCMCs)”)
First known “language model” - 1913: First 20,000 characters of Eugene Onegin



Image


Slide 4

Image


Slide 5

Image (Apologies for English language bias!) Let’s start with the letter ‘s’. What follows? (Can someone keep count?)


Slide 6

Image

Image


Slide 7

p h t s k . SPACE [e…] s 1 1 1 1 1 1 1 0… s 1/28 1/28 1/28 1/28 1/28 1/28 1/28 0/28…

Image


Slide 8

Image


Slide 9

And just like magic…

(After we download Eugene Onegin in English and convert it to plain text format)


Slide 10

Image


Slide 11

Image Image

Image


Slide 12

Markov (Language) Model

Models independent probabilities (“z given y, given x, given z” etc etc)
“If I have a vowel, what should come next?”
Roll a die / clip a coin
16% - another vowel
84% - another consonant

Image


Slide 13

Are Markov Chains Large Language Models?

Tokens (sub-words) not letters
Conditional probabilities: not one but thousands, millions…
Multidimensional (‘s’ relates to ‘e’ on dimension X, not on dimension Y)
Positional encoding (where a token lies in a sentence)
“Transformer”: architecture that looks backward and forward (“attention” mechanism) to transform input into output
But still… basically very fancy Markov chains
What does the “expert” (GPT) say?



Image


Slide 14

Let’s build a Giant Markov Chain. Let’s call it a Generative Pre-trained Transformer (GPT for short).

And we’ll use the whole Internet (instead of just Eugene Onegin).

And lots of data and GPUs (which are good at matrix multiplication - manipulating spreadsheets).


Slide 15

What Could Go Wrong?


Slide 16

What is the Internet?

Eugene Onegin
Wikipedia
Millions of books
Billions of webpages
Social media (Facebook, Reddit, Twitter, Tiktok, YouTube etc)
Most content is unmoderated, can represent overt (toxic, offensive) or covert (non-obvious, unconscious) bias


Slide 17

Bias

Image



Pre-GPT models had problems with bias
See classic study by Bolukbasi et al. (2016)
Bias can be quantified and modified
But Transformers (Vaswani et al. 2017) are so large and complex, debiasing is now much harder


Slide 18

Common forms of bias

- Gender
- Race
- Religion
- Disability
- Sexuality
- Class
- Other social markers

- Bias can often be intersectional (Crenshaw 1989) - involving more than one category
- Do LLMs exhibit intersectional bias?


Slide 19

Intersectional Bias in Causal Language Models

Image

- Tested on GPT-2
- Categories of disability, gender, religion
Main findings:
Reproduced bias towards Muslim religion, physical disability
Associations of violence, terrorism, criminality
Intersectional bias is not always additive (X + Y). This means LLMs cannot easily be debiased.


Slide 20

Alignment

- Tested with GPT-3 onwards
- Humans score LLM outputs
- LLM is re-trained on these scores, and adapts its weights to produce “better” outputs
- Important paper: Ouyang et al. (2022).
- Make GPT-3 more helpful, truthful and harmless
- Debiasing through making LLMs outputs correspond to human values (preferences, desires - and also biases)
** - **Breakthrough that led to ChatGPT in November 2023. Largely (not fully) addressed issues of bias.


Slide 21

Alignment

- More questions: alignment with who? With what?
- Alignment has become the key “problem” that organises vast capital ($100s of billions today)
- What is helpful? What is true? What is harm?
- Computer scientists have become parents, therapists, teachers for LLMs
- Interesting history: research by the Moscow Linguistic School (1950s, 60s) - how to reconcile language structures (or models) with statistics. These problems are old and not resolved…

Image


Slide 22

Relevance for Teaching (Finally!) Image

- Algorithmic bias can reproduce social bias
And limit learning for students from diverse or minority backgrounds - Can bias ever be eliminated? (and who determines bias?)
- And alignment can introduce new biases that affect how students learn
- Do we need AI literacy, to help us interpret bias, alignment ideology? (see Critical AI)