[A PROPOSAL FOR A PLAY]

RENEE: We are reading through the data sets.
SONIA: ISO is the standard verification language. It is the the standard association that everybody guides themselves by in the technical world.
RENEE: ISO Standards from pizza crusts to education to AI. It is quite something.
SONIA: reference ISO gym
DONNA SUMMER: I want to look at the code! The actual code and AI. It has a supernatural quality that I would like to incorporate. Example: Tesla car near a cemetary. The car was detecting shapes and energy but there was nothing there. Did AI detect a ghost?
AGATHE: I have code for how to detect objects.
DONNA SUMMER: Information and context. How does that work as well?
RENEE: oOw does the camera frame the recognition?
SONIA: Let's look through several introductions.
MICHELLE: I am doing live transcription and intend to write a play from this meeting. Would it be possible to have the screen share on so I can look at the code while I am writing?

[brings up Microsoft COCO: Common Objects in Context]

A: Let's look through this link, then the pdf.

[opens dataset tab]

S: do you see it?
A: here is a summary of the data set s... persons, people, objects. They show what is annotated under the images. Pictures with computers is important to annotate.

[cats and computers, sinks, floors, cats in bottles, cats on a chair, cat on a sofa]

S: Our future is imagined through these flat surfaces instead of 3D spaces (with sounds)
R: COCO comes from the Gorilla, Like COCO wants a banana?
S: Perhaps happy coincidence.
S: Should I click on something else?

[brings up horses, in landscapes, hills, towns, etc]

A: These have been critized because of these connotations.
S: Of course it is going to be pretty safe. The scandal with Google. Image recognition with people recognized as animals, and...
A: Pictures ... recently of a hand holding a thermomator .. Concentration camps as leisure parks
N: Interesting the cell phone, and the icon. How are the icons developed?
A: Interesting to see what they decide to annotate ... home images... subjective. The whole person, the face of the person. There isn't a lot of reasoning behind these piectures/ They seem quite random.
R: It is scraping any informatino from the url> ?? How much is being read. Whether that url is adding, framing the dataset. if you have the url there will be textual framing.

[guy with home cooking blog]

R: The circle is a cake.
A: Most datasets scrape images from the web. Generally they don't save any context information. They send the image on a cross reference platform. They ask precise questions to annotate what is in the dataset. Or precise segments. Either they force a certain set of annotations. Or let the workers annotate whatever is important in the image. There were pictures of people annotated with character traits. ... associate and full professor. You find a lot of hierarchical things in the dataset.
C: On the interface. They create a lot of interfaces. they are built for speed for speedy workers. Then another richer slower interaction interface. The level of communication between these interfaces was very interesting. Categories: the amount of sports are quite limited. But there are two for baseball and the baseball glove. Pizza, food, child's food. Reading the paper, the three sources for categories. Children 8-14 asked what is the most prominant feature in their environment.
A: They said they wanted to capture their whole world. And asked people to take a picture in their environment. So it is very telling.
L: Why children?
C: Assumption that children as the least biased. The difference between humans learning and machine learning.
A: To augment the data set they asked children to do it.

[showing pictures in front of computers, outside, on sidewalks, in diners]

A: Pictures of weddings. White long dress.

[Normative, Westernized images]

R: Mechanical Turks and where they might reside. There has to be a contradiction.

[Army photos, people on phones, playing baseball, chopping onions, holding a cell phone, computer sccreen]

R: It raises the question who is being trained actually?
A: The workers train themselves to have the same strategy.

[Who and where are the workers?]

A: A soup is generally in a bowl. Workers are flirting.

[laps, at a table, lawn chair]

R: To capture the whole world is almost an art project like Kenneth Goldsmith printing the internet.
A: Should we read the paper or look at something else?
A: The workers are mostly coming Amazon's Mechanical Turk.

[Jeff Bezos refers to these workers as AI]

S: The more you work like machines the more you become one.
S: The background never matters. It is only about the object. But the context is makes the object. \

[Context matters]

S: Cars in the background. Even if tiny are still labelled as something to be presented.

[https://unthinking.photography/articles/an-introduction-to-image-datasets]

R: It is so interesting thinking about what context is. How much photography theory we had to read. What is happening inside and outside of the iamges. What happens in the web context where images can land in so many different contexts.
S: Segmentation. 2,500,000. 22 worker hours per 1,000 segmentations. I am trying to do the math. This is a lot of time! The need to do one task, super fast.
A: Having one person who has one task. And becomes super familiar with that task.


[scanning through several graphs]
[Fig 6: Samples of annotated images in COCO dataset]
[Person Bike Horse]

S: So COCO is better. Jesus 7,000 workers hours.
A: They evaluate training of datasets.

[Object & Scene Categories]

DS: I really like the captions on the categories.
DS: A group of human judges evaluating whether the caption is right or wrong. There is something poetic about these captions. Woman in dress standing among them. Something literally. A person dressed in colourful clothes. \
R: Almost like Haiku
S: The situation of being in front of the computer. Recognizing an object as fast as possible. Is so alien of the situation of walking around on the street. The way to very forcefully introduce this information into these machines for events in the future is completely at odds of the experience of walking throught the city. Perhaps a person wearing Google glasses and annotating the experience of walking would be more interesting.
DS: Sounds like an over-estimation of reality. Most people don't walk thorugh the city that way. I mention the example of the captions because it reminded me of the experience of creative writing. Good stories are demonstrating something in an objective way. And it is you giving meaning. Chekov writing is not too different from the captions of COCO. We also always try to be purely descriptive, which is hard when today everybody is trying to giving opinions, and advancing their objectives. As researchers, we always try to spot the moment when the machine has the bias. We already know the answers, that machines are going to be racist and sexist, which are valid points. But it is not enough of this approach that finds something in this that can advance the arts, fine arts and scholarship of what these things can provide.

[In this world, umbrellas should be in toilets, and cats in sinks]

R: So they are purely descriptive as opposed to interpretative.

[Group fairness is the goal of groups defined by protected attributes receiving similar treatments or outcomes.]

[Individual fairness is the goal of similar individuals receiving similar treatments or outcomes.]

C: The glance vs the gaze. The workers of Imagenet. Doing it in person, writing the notes. One person gets three images. Then a group of three coming up with a common description. Glance: we only had a few seconds to look at the images. After we became more accustomed to looking at the images. Researchers did that to avoid bias. If you don't have time to register what you are looking at, then there is less bias. Going through this, and collectively was an amazing experience.

[Bias is a systematic error.]

[https://experiments.runwayml.com/generative_engine/]
[http://moments.csail.mit.edu/]

[https://www.google.com/url?sa=i&url=https%3A%2F%2Fwww.groundai.com%2Fproject%2Fdo-imagenet-classifiers-generalize-to-imagenet%2F1&psig=AOvVaw1MKV_xTVY6u04H8kRsEfOc&ust=1615992357643000&source=images&cd=vfe&ved=0CAIQjRxqFwoTCPCYjeyGte8CFQAAAAAdAAAAABAJ]

A: This is what the ImageNet crowdsourcing task looks like

S: I have to leave to teach at 4:00

R: The glance and the gaze. The wink and the blink. We all blink. The wink is intentional. Funny to see that language echo.

R: On terminology. The desire to use more neutral language.
R: Interesting in that text, demonstration and not explanation. The cold eye, what would it do... 15 years ago I had a Polish student in PZI who was interested in political media images. He took super violent images and described them clinically. White and black drawing. No connotations. Violent images but absolutely neutral. And interesting in this case where bias is a computational error.
A: Lighter skin coloured people. So Twitter started to say that it was not a problem of fairness. That it didn't need to be systematic to talk about fairness.
R: And what if they were to say "bias is a sin".

[reading]

A: Defining a protective and non-protective group. Detecting bias. You should look at fairness by looking at attributes. But now that attributes are not universal...
R: And they are also trying to design themselves out of a problem. Designing out of bias.
A: The fairness matrix to determine whether something is fair or not. It is a bit of a design problem because you have to assign the attributes. But the matrix and methods have a lot of problems.
A: I still cannot share my screen.
R: The language of 350 degrees. There is something inbuilt in that title because it says ... it is weirdly acclaimed to universal knowledge. Instead of how do you make bias transparent it is expunging. 360 degrees is an interesting starting point.

[obliterate or remove completely something unwanted or unpleasant]

R: Training models without biases. Is this where we are at?
A: Making a system to predict whether somebody will have to go to the hospital or not. To predict whether there will be expenses. There was a dataset on people who had been to the hospital for long periods of time to develop a dataset based on fairness.
A: This whole section. I am not sure how to go about looking at the code.
R: I love this prejudice remover. Glad they have an algorithm for that.
A: I have never read code with people. Not sure how to go about it.

[This is the starting point]

A: Now they have their data set and they are going to predict their model.
R: It is great to see 'privileged groups', and 'unprivileged groups' to get an idea of semantics.
A: Mostly they are picking and defining values based on race or gender or a combination of the two.

[scanning through code written in Python]

Protected attribute names
('RACE;)

Cannot do more than two groups at once (0 and 1) binary

Privileged and unprivileged protected attribute values

3.2.2. Validating LR model on original data

unprivileged groups=unprivileged groups,
privileged groups=privileged groups

print(explainer_orig_panel19_train.disparate_image))

from collections import defaultdict.

def text(dataset, model, thresh_arr)

C: Do you think they developed it themselves?
A: This validation model. Each is aligned to one fairness metric. I find it interesting to define the metric you have to define the protected group. So it simplifies the whole context.
R: Where things go array... the arc of the process is pattern recognition and extrapolating conclusions. Bias can come in at any of those junctures.

[Classification Thresholds]

A: They don't just have a binary decision. Choosing between the threshold between 0 and 1 to determine whether a person should go to prison for example. The red line is the fairness. Practitioners need to make a choice between the threshold.
R: So this is almost a Mechanical Turk moment. Where a human says, a little tweak this way, a little tweak the other way.

[But what do these numbers mean and how do we apply them in practice]

L: It is interesting that you are trying to mitigate bias in the program but in the end you have to use bias to make a decisions.
L: Using the language of mathematics to describe an emotion. An inappropriateness of language.

[Teknik and Magic]

A: Fairness based on their own metrics.
A: Issues of discrimination come from bias.

[Threshold corresponding to Best balanced accuracy: 0.1900]

R: I don't want to hold engineers to a higher moral standard than myself.
R: We are in an interdisciplinary moment where we need each other to think through the consequences. Looking through these documents, I need to see the logic. How to engage with the language with understanding the limits. (regarding education).
A: Regarding education... Mostly these conversations (discussion the socio-technical limitations) do not take place.
R: Arts have limits as well.
DS: This reminded of my classes on distance reading. In the humanities, many people use this methods to establish patterns between huge amounts of text. Critics of the methods of distance reading. In literature we are not interested in patterns, but exceptions. That is where the interesting story exists. Not trying to solve the bias of AI. It will exist nonetheless, but use this moment, these glitches of the code for a space of new knowledge, new art, new generation.
DS: After years and activism, I am trying to use a new lens of fun to look at things.
L: Not trying to solve bias through code.

[Or use those biases as a source of debate and public engagement?]

L: Playing cards. Playing the dataset. You can see algorithms as divination. Maybe bringing it out of the rationality.

['Algorithms as Cartomancy', Flavia Dzodan]

[An escape game to make people aware of AI bias https://www.tomokihara.com/en/escape-the-smart-city.html]

L: We need to take it out into another type of field. Or looking at it differently.

[END]