Combined

Kaka-Problem

What to find out how many are in wellington, how do they move around but they are mostly unbanded

Kaka- Input

Pictures of head and beak, as they have distinctive beaks

Kaka-Output

Kaka- What features on beak

• Beak shape: curvature, width/height ratio, length

• Beak texture: gradients, spots, imperfections

kaka- Problems

How static are these features

data collection with kaka notoriously curious

Kaka-Output

Re-identification or identification, New or not

Kaka- Model

Unsupervised

• Image segmentation

• Feature extraction and matching

• SIFT

Deep Learning 

DINOv2 (a vision transformer model)

Modeling Plant Responses- Problem

• Ecosystems are changing...we don’t fully know why

• Species move or go extinct: range shift

• Can we predict where a species is likely to move?

• Can we understand the ML model to see why?

Plant-Different Models

Random forest, Logistic Regression

all have tradeoffs

using KNN to calculate noise level and remove bad data

Healthcare-Problem

Limit Staff

Underfunding

SkinC -Input

Pictures of Moles

SkinC - Output

Benign or malignant

SkinC- Model

Classifications

Healthcare Potential Issues

Data needs to be representative of whole population

Needs to be medical grade

Te Tiriti O Waitangi

Data authority, ownership, confidentiality-opt in?? be good

Oversight actually needs to be supervised

Vision of AI

magine a doctor’s practice in 2035:

We can imagine various AI-based labour-saving devices being

used by doctors. . .

Likewise for other healthcare workers. (Nurses, administrators. . . )

We can also imagine various AI devices/systems being used by

patients, before & after their appointments.

Some healthcare consultations can be carried out ‘at home’.

We can also imagine various AI systems for making hospitals run more

efficiently. For instance:

Triage & prioritisation

Surgical planning

Predicting the ‘flow’ of patients in and out of hospital beds. . .

Predictive AI systems

operate in a specific domain.

- They learn to map inputs in a given task onto outputs, using

supervised learning.

- They are relatively easy to evaluate.

E.g. you can measure accuracy on held-out training examples

Generative AI systems

They learn to generate content of a given type.

E.g. text, images.

- Their power comes from training on huge datasets of examples.

- Text generation systems (e.g. GPT) are the focus for medicine.

Ali Knott AIML131 Week 9 Lecture 2 7 / 23

Other uses of AI in healthcare

Predictive AI systems are the dominant paradigm here.

E.g. a system that identifies fractures in an X-ray image

E.g. a system identifying white blood cells in microscope images.

Weather Forecasting-Problem

Expensive,

Inaccurate in NZ - terrain, roaring40s, isolated

Computationally Expensive 

Weather Forecasting-Input

t

Weather Forecasting-Output

t+1

Weather Forecasting-Outcome

cheaper, and more computationally efficient

Weather Forecasting-Issues

earths a sphere 

Climate Change

If AI models are trained on past weather forecasts,

how can we be sure they will accurately predict

the future weather under a warming climate?

• Many AI models focus on only predicting the next 10

days, so if the initial state is warmer, the forecasts will

be too

• Continuously re-training (AIFS does this)

• Foundation models built on climate and weather

information

• Ensemble-based models

• Post-processing for extreme weather events

• Including climate forcings (CO2 ppm, sea surface

temperature) as a predictor

Weather Forecasting-Models

NN and Auto Regression

CNN and MLP

Transformers

Other AI applications of Ai in weather

Cyclone Tracking

Flood Forecasting

Sea Ice forecasting

Recommender systems-def

push content at users

Harmful content classifiers-def

that withhold content from users

Withholding content- things to consider

What classifies as harmful and who should

-classification done by humans but too much so Ai as well

Ranking Harmfulness→ supervise learning classifer

Remove, does nothing,borderline→ downrank

Censorship-Bias

Transparency is needed

Harms of AI in SM

keyhole, as it recommends content users they lie, pushing to radicalization,

mental health- showing perfect life

Spread of hate speech Andrew Tate

Christchurch call -> eliminating TVEC

asking companies to remove content of intimate photos. Also penalties for rumors and threats

AB tests

A/B tests: create random user groups, and give them different

versions of the system. Then measure their behaviours.

We asked companies to do A/B tests to test if they could change

the amount of TVEC users saw. They all said no

EU risk levels

Unacceptable, high and limited risk

EU unacceptable

Systems using subliminal techniques, purposeful

manipulation/deception

Systems exploiting vulnerabilities of individuals or groups

Systems identifying individuals through biometrics (e.g. faces)

- With exceptions for law enforcement

Systems for ‘social scoring’

Systems for inferring emotions, in workplaces and educational

settings.

EU high risk 

safety (medical)or social issues (bias, employment, justice)

AI Act provisions about Gen AI 

Publish summaries of the content 

Copy with EU copyright

Write technical documentation for ‘downstream users’

The Bletchley Park AI Safety Summit

Cooperation with over 28 countries NZ joined in 2024

This event focused on large Gen AI models

recognizes risk of bio weapons, toxins, alteration of genes and cybersecurity

-transparency, evolution metric, safety testing

Recognises the risks of ‘frontier’ AI models

Commits signatories ‘to work together in an inclusive manner to

ensure human-centric, trustworthy and responsible AI that is safe,

and supports the good of all through existing international fora and

other relevant initiatives, to promote cooperation to address the

broad range of risks posed by AI’

UK legisatation

Still a year away

security institute, hand over model to be tested, copyright 

US Policy

Joe biden had an order to disclose risks to national security results

develop standards for red teaming and evaluation

Donald trump removed that replaced it with how to remove ai barriers

China

No comprehensive policy’s, but big reports & committed

detailed legal requirements on AI content labelling and

watermarking are already in place

-thinking about cybersecurity, biosecurity and open source

Job impact

number of young employees declines

productivity

optimus humanoid in manufacturing, or self driving delivery vehicles

Managers - gig work

Issues in Ai in workplace gig work

monitoring

Recruiters use of Ai

In CV and Cover letter screening

Ai interviews

Ai taking notes

employee/candidate use of Ai

Using Ai for Cover letters and CV

help in practice interviews

and ai in workforce

1 Scenario of AI in NZ

The main effect of AI is to improve the productivity of workers.

No mass unemployment; instead, NZ workers are more efficient

Positive uses of Ai in recruitment

Positive uses of AI in recruitment

We can also work to build fairness into recruitment tools.

1. We can delete features that aren’t relevant from training sets.

Gender, ethnicity are often irrelevant. . .

If assessors classify applicants based on redacted application

materials, it’s harder for them to be biased.

2. We could include audit functions, that show percentages of hired

people from different demographic groups.

This way, biases will at least be visible. (Within the company and

beyond.)

Working alongside AI

Say you’re a worker making decisions. . .

- Perhaps you’re a doctor, looking at X-rays and detecting

fractures. . .

Say an AI system is working alongside you, to help you.

- Say it’s pretty reliable. . . perhaps 95%. . .

How do you stay in the loop??

- It’s hard to stay in control, if the system works well!

Also - who’s responsible, if the thing you are jointly doing goes wrong?

Ali Knott AIML131 Week 11 Lecture 2 17 / 24

Scenario 2

Replacing onshore

NZ workers are displaced by AI, into lower-value work.

The AI systems doing the displacing are NZ-owned.

(So NZ can recover some of their profits through taxation.

Replacing offshore’

NZ workers are displaced by AI, into lower-value work.

The AI systems doing the displacing are owned offshore.

-International tech tax

International Tech Tax

There is an international tech tax being organised by the OECD, for

large multinational companies. The US walked out in June.

‘Pillar 1’ sets things up so each country taxes LMCs according to

the revenue they make in that country.

‘Pillar 2’ establishes a ‘global minimum tax’ of 15%, so companies

can’t run to countries with low corporation tax.

GAN’s

Generative Adversarial Networks (GANs) arrived in

2014, dominated many fields until 2020

• Train two models in an adversarial fashion and they compete with each other

• Generator: E.g., a CNN model producing candidate images

• Discriminator: E.g., a CNN that judges if an image is real or

fake

• Training two NNs at a time often leads to instability

CNN for image generation

Popular in 2012

Not generative, but form backbone of GANs and autoencoders

Autoencoders

Appeared in 1990 revived in 2013.

encoder compress an image into a smaller, latent representation using a NN/CNN

Decoder learn to deconstruct the original image from latent representation

Diffusion 

Gradually add noise to an image, then train a model to

learn how to remove the noise step by step, to get back

to original image

• At inference, start from noise, denoise step-by-step

• More stable than GANs and high quality images, but

slower to train and use at inference

CLIP

Contrastive Language-Image Pre-training. Maps pairs. matches pairs using the cosine vector equation. encodes to the same shared space. It is a model from OpenAI trained to link text and images through a

shared encoding. Open AI in 2021

Forward / noising process:

goes from a structured image→ noise

Reverse/denoising process

trains a NN to learn how to remove the noise step by step

3 steps for image generation Diffusion

1. Encoder Text embedding-converts a text caption to a text encoding (a

   numerical representation)

2. Prior predict what the image imbedding will be. eg CLIP

3. Decode- converts the image encoding to an image

What does ChatGPT use?

Autoregressive multi-modal model. NOT DIFFUSION

Text-to-Video

Generates images from text using diffusion then uses image interpolation

Video Generation Models

Sora by OpenAI

Veo3 by Google Deepmind

Other AI image tools

• Upscaling

Inpainting -Filling in missing parts of

 an image, or generate

things that aren’t really

there

• Outpainting- used to generate wide shot shots, extending an image beyond its original boundaries

• Captioning

AI Art Ethics

Using Ai to make images - competion entry

What is art?

How do we copyright? - A comic book written entirely by midjourney - copyright was revoked as it only protects human created work, however the arranagment and story is copyrighted

Music- music not copyrighted

Harms of Generated Images

Bias - beautiful fair skined, feminine, smiling, soft emotions

Cultural appropriation

Image ownership for trainning

Benefits of Ai generated images

• MRI Denoising with Diffusion Models

• Improve diagnostic tools and patient outcomes

• Enlarging datasets of rare conditions, upscaling low res scans,

• easier prototyping and iterating for architects -cheaper and quicker than human building digital photos

• AI generated scenes of hazardous driving conditions for Ai to learn from

DeepFakes

an image or recording that has been

convincingly altered and manipulated to

misrepresent someone as doing or saying

something that was not actually done or said

History of Deepfakes

Orginally using GANs 

Now can use a range of different models

• diffusion

• transformers

• autoencoders

GANs

train the discriminator to classify which image is fake or not

train the generator to generate images from random noise- using different models such as diffusion or an autoencoder

Penalize the generator if the discriminator can tell the difference, use backpropagation to update the weights until the discriminator cannot tell the difference and just guessing randomly(50%) 

Deep Fakes used for bad

Political examples

- gabonese president was absent from public due to medical situation and a deepfake was sued to show he was ok, however it was suspected of being a deepfake

-a video of Ukrainain president telling his army to surrender

-Joe bidden not to vote

Celebrity

-scams people thinking they are in a relationship

-Promoting products unaware

Pornogarphy

celebrities targeted

Women in politics

Young girls, teachers

Finace

Cybersecurity -scams, and used to bypass biometric privacy

Regulation in Politics

June 2023: The European Union’s EU Act requires AI content to be

labelled, especially in political contexts

• Jan 2024: South Korea prohibited the production of political

deepfakes for the 90 days leading up to national elections, unless

they were clearly labelled as fake

• April 2025: Singapore ban deepfakes misrepresenting political

candidates more

• American state-by-state laws:

• Texas: criminalised the creation and distribution of deepfakes intended to

influence elections within 30 days of the election

• California: bans publishing materially deceptive deepfakes of politicians

within 60 days of an election, unless labelled as fake

Good of Deepfakes and AI video

• Translate in different languages, in sign language, allows people with motor disabilities to express themselves with a synthetic voice.

Social Media

X thinks good, humour, censoring

Meta- Ai info label for users to indicate, try but numerous test show that their scanning doesn’t work that effectively

Tik Tok - Asks users to self disclose, can label some content as AI generated

Detecting Deefakes

Harder and harder for humans.

Intel’s FakeCatcher claims a 96% accuracy rate (in milliseconds)

• Looks for subtle “blood flow” in video pixels!

• Our veins change colour as our heartbeats

• Social Media AI detectors require the videos/images to be digitally

watermarked by the AI deepfake models

For AI-based detection (like those used by social media platforms)

they essentially need to build a better discriminator than the one

used in the original GAN

Challenges of image generation

Realistic conditions: Lighting, shadows, and reflections

• Minute details: wrinkles, water droplets, hair, fingers!

• Image-wide coherence: everything in the image needs

to fit together, scales need to be correct

What is Māori Data Sovereignty

Māori data is data produced by Māori or about Māori

or the environments that Māori have relationships

with. Data is a living tāonga and is of value to Māori

example of Māori Data

Data from government agencies, organisations and businesses

• Data about Māori used to describe/compare Māori collectives

• Data about Te Ao Māori that emerges from research

What is Māori Algorithmic Sovereignty

Algorithms that use Māori data, or

are applied to Māori individuals, collectives, or

environments that Māori have rights or interests in,

should be subject to Māori governance structures

Rangatiratanga

Authority

Control

-permission, how it is collected, used and stored

Jurisdiction should be stored/deployed in a way that enhances control, within Aotearoa

Example: using a NZ cloud service (e.g. Catalyst)

where people have opted-in (informed consent) to

their health data being analysed

Whakapapa

Relationships

Transparency- whos involved, deployment, explainability

Data Relationship- how is Maori data used throughout the algorithm

Sustaianability

algorithmic outputs are used for long-

term sustainable benefit to Māori & environment

An indigenous example: the “Pima Indians” dataset was

collected from a long-term study of indigenous peoples

in the USA. Became used as an ML “benchmark

dataset” without any real permission from the Pima

Whanaungatanga

Obligations

-balancing rights and accountabijlity

Kotahitanga

Collective Benefit

-output in future helps Māori or the environment

-Builds capacity, development of a Māori workforce –

empower Māori and learn by teaching (ako)

• Solidarity: supporting connections between Māori

and other indigenous peoples to enable sharing of

ideas, strategies for fair algorithmic development

Manaakitanga

Reciprocity

respect

consent

privacy

Kaitiakitanga

Guardianship

Ethics

Restrictions- Māori should decide if input/output is tapu or noa

Overall 6 groups

—

Why XAI is important?

Insights, trust, Clever Hans, detecting bias, saftey