 Thanks again for joining. I know that it's busy and taking time out to attend a mock class is yet one more thing in a very busy schedule. But what I really want to do with you tonight is to introduce you to the world of image classification and some of the ethical issues that have infected, and I use that word intentionally, infected the search results that you get when you use tools like Google Image. So by the end of our time together, my goal is for you to have a better understanding of how image classification actually works conceptually, anyway, not technically, but more conceptually, and some of the flaws in the early image classification work that still affects search engine results and how and who people are represented. All right, so, okay, good. Looks like everything's working. Give a shout, Lily, if anything doesn't look right. So, most of you probably use Google Image Search. You use it for a variety of reasons from searching for favorite musicians or rap artists, maybe answering homework questions. Maybe you've had to put together a presentation for an earth science class and maybe on the solar system and you get research results like this when you search for solar system and it allows you to get that perfect photo or diagram to capture the attention of your audience, like cute puppies. But have you ever actually really stopped and looked at what Google actually retrieves for you when you search for images and why you get some of the images that you do? So I'm gonna give you an example. Yesterday I went into Google's Image Search and I typed professor. This is the first screen of images that I get when I do that search. So my question to you is, what do you notice? And again, feel free to type in your thoughts or raise your hand and I will pay attention to the chat window. In fact, let me move, I have to move some images around here, too many little boxes on my screen. The money heist character, yes. What else do you see? Again, the category here is professor. There's a lot of men. Any other observations? They're all white, good. Wikipedia is the first result. Actually, it's a good observation. So a lot of the images coming from Wikipedia. You have chalkboards. Actually Ben, one of the things that struck me is apparently a lot of professors teach math because the chalkboards almost all of them have formulas on them. We don't teach anything else. Good, okay. So what caught my attention as you guys noted is that it's largely men. There's a couple women in the image. And I got wondering if maybe if I search for young professor, if that might give me a different result. Maybe older professor, older professors maybe tend to be male and white. So a search for young professor with maybe the professoriate becoming more diverse would return a more diverse search result. Here's what I got when I searched for young professor. What do you see? Again, feel free to type your observations here in the chat. So mostly men, white men. Yep, exactly. I think again we've got what two women in this photo. It's stock images, 100%. And so if you start drilling into these, you'll see that they're stock images. But that in itself doesn't really matter because it's about this categorization that's happening of the category of professor. And so that's what I want us to think a little bit about. So the question I'm going to talk through with you a bit today is why? Why is it that when you are doing these search results, especially around people, you get the kinds of results like what I just showed. So one thing that's important to keep in mind is that search engines, the algorithms that drive search engines, they are very powerful and of course they're pretty helpful, but they are also flawed and especially when it comes to categorizing humans. So I'm going to give you a quick crash course on what it means to classify images because when you search for an image, you're entering a word or a set of words that define a category that is in driving that search for images that match that category. So when you look at this image, a word likely pops into your head. What is it? Feel free to type it. I see dog, yep, I see puppy. Lots of puppy, lots of dog. Oh, winter dog. So the label of dog or I'll stick with dog. So the label of dog is an artifact of humans who speak English in this case over time using and agreeing and reproducing that sound dog and the spelling DOG that's associated with that object and of course variations on that object, right? We recognize as humans that we humans are able to recognize that all of those are part of the same category of dog. So let me give you a couple of other images. What would you say the category or the, yeah, what's the category of this image? What label would you put to it? Building, old brick building, good. Any other thoughts? Oh, architecture, interestingly. Okay, let me give you another image. What would you say here? Bird, blue bird, right? Good, all right, so I'm gonna give you another image. How would you classify this? Oh, Arizona, nature, environment. Interesting that you guys go to states. Oh, landmark, right? Yeah, erosion, rock formation. Yeah, Grand Canyon Park. I'm actually not entirely sure where this is. Any other labels or words that come to mind when you think about this one? Rock formation, that's one that came to my head when I first looked at this too. So this one is harder, right? So you see the variety. When you guys did, when I showed you the building, you just had building for the most part. Bird was bird, although I noticed some of you added blue. So it was blue and bird, two different words. So you have these main categories. In this case, maybe rock or rock formation, but then we tend to assign attributes to those images like in this case for this photo, maybe red or beautiful or even something like Arizona or dry are attributes of the object. Because the object itself is not Arizona, right? Okay, so how about this image? What's the first word that pops into your mind when you see this? Woman, black woman, good. Any other categories? Professional, office, anything else. Confident, right? That's sort of this, good. So, independent woman, excellent, good. So when we categorize, as I said, we're not only looking at this image and giving it sort of this main category. In this case, the main category might be mammal, actually, or person, but then there's all of these additional attributes like black or professional or independent. Even woman is an attribute of person because it labels their gender. So naming exerts power on that which is named and classification is naming and therefore naming is power. That is a syllogism and it'll be a bit of a theme here as I work you through the next part of the conversation. In order to work you through image search and image classification, I need to do a very quick kind of crash course on how image search works. In other words, how image classification works. Computationally, that is. Okay, so hopefully this isn't too fuzzy. I grabbed this screenshot. I went to Google Image, actually, and I found this diagram of image classification. So the idea here with image classification to be able to then do something like what Google does in giving you search returns for, in this case, dog or cat, I'm gonna walk you through the process. So we're gonna start over here with labeled data. So if I wanna categorize images of dogs and cats, I am going to have you guys and maybe a bunch of other people look at images of dogs and cats and frogs and other sorts of images and have you label those so that you basically have a large data set of labeled images. And then I'm also, of course, going to have the labels. So I need the machine learning algorithm that I'm training also understand what the categories actually are. So I'm gonna say cat, dog and frog, and I'm gonna feed that in to my classifier. And again, I'm not going to unpack convolutional neural networks or any of the approaches that go into developing the algorithms. But just trust me when I say that if you feed in lots of labeled data, then you can get the software to begin to see the patterns in the images that are labeled for dog and those different shapes and colors that our brains do a very good job at when we look at different images of dogs and say those are all dogs. Over time, you can get the computer to also be able to recognize as long as you have lots of labeled dogs that these different shapes and features are dog. And eventually you get a classifier that can accurately differentiate a dog from a frog and a cat from a dog. That's the power of image classification. And this is a, I don't know, a 40 year enterprise that has been undertaken in computer science to get to the place where you can go into Google image and search for dog and get lots of images of something that you would recognize as actually being a dog. Not an easy feat. Took a lot of time, a lot of hard work. And that's what I wanna talk about next. So in order to create the artificial intelligence, the algorithms that categorize images, you need a large dataset of already labeled images that are used to then train that model. So I'm gonna tell you a story about ImageNet, which is, or was, it still is, the first of its kind large dataset for labeling images. So in fact, you can go to the ImageNet website, the URL there is at the top. So ImageNet was the brainchild of computer scientist Dr. Fei-Fei Li. This was back in the mid 2000s, so 2002, 2003. She was an assistant professor in computer science at the University of Illinois, Urbana-Champaign. And in those earlier approaches to image classification, it wasn't very good. The algorithms were not accurate at identifying images correctly. And she observed that one of the problems is that there wasn't enough categorized training data that would then help get the computer to recognize that a dog is a dog in an image. So she came up with the idea of providing a library of categorized images to train the software to then help build these algorithms. She actually scraped hundreds of thousands of images from Google, not Google image, but just searching Google, going to websites and pulling these images. And she then hired thousands of people through something called Amazon Mechanical Turk. And I'm not gonna talk about EMT right now if you have any questions, I'm happy to talk about it. But she basically hired people to look at thousands of images. And in the end, she was able to categorize 12 million images, it's a lot of images, with 22,000 categories. And then she released that data set in 2009 for companies and researchers to train their algorithms to do a better job of predicting or identifying objects and images. So, but I mentioned 22,000 categories. That's a lot of categories. And so where did all those categories come from? Well, they came from WordNet. So WordNet was developed by computer scientists at Princeton in the 1980s to aid in textual classification work. And again, you can go to the Princeton website. I must confess, I didn't look to see if it's still up, but it was, I don't know, it was a year ago. So the way that WordNet works is it categorizes words by their main categories, by their subcategories, and by the attributes of those subcategories. So I'm giving you an example here of the image is three sheep being chased by a dog. And so if you were to categorize these, you've got the dog and the sheep are mammals. So they're under the mammal category. And then under dog, you've got border collie and different attributes like loyal and hardworking. And then under sheep, you have merino and herbivore. And so that's how WordNet is structured. So what's the problem? Well, the problem is that WordNet has a people category. And that people category has a number of subcategories and then attributes of those subcategories in it. So I have highlighted here, sensualist. So sensualist is a subcategory under person. And under sensualist, there's a subcategory of bisexual. But categorizing, just think about this for a second. Categorizing bisexual under sensualist, but not including heterosexual or homosexual, basically positions bisexuals as somehow more, I don't know, sexualized than people who are straight or people who are gay. And it also conflates bisexuality, if you look below in the subcategories there, with, so bisexuality, which is a sexual orientation with hermaphrodite. Hermaphrodite's a biological condition where a person has both sexually reproducing organs from, has sexual reproducing organs from both sexes. So, and I must confess, as someone who is bisexual, I'm kind of offended by this, but it doesn't stop there. So under the person category, there also were several other kinds of problematic labels, including loser and dud, abandoned person, I don't even know what that means, partner, beggar woman, again, you can kind of see the list there. Anything from loser to underdog and all kinds of things in between. So, a decade after ImageNet was released, again, ImageNet was released in 2009. And two AI researchers set out to highlight the problems with ImageNet's people category. So they created a website called ImageRoulette. So it was part activism, part research, and people could basically upload their photos or photos of other people and get a sense of how ImageNet was categorizing those people. Unfortunately, the website doesn't exist anymore. It got taken down about a year after they put it up. But thanks to the internet and the fact that the internet never forgets, there are lots of screenshots about some of the different ways that ImageNet was categorizing people. Let's take a look, shall we? All right, I'm just gonna give you a second to absorb these. What do you see? Feel free to type in any thoughts or observations in the chat. Adjectives. Do these adjectives make sense for these images? Camila says highly derogatory terms. Are there any in particular that strike you as problematic? Oh, Amy says subjective terms, that's good. Right, so Janth says the person on the right is described as sick, but they don't really look sick. William says, yeah, stereotyping without facts. It's interesting that you mentioned stereotyping because from a 10,000-foot conceptual idea of how classification work, it's about stereotyping. You're trying to think about what the set of similar attributes fit a category. And stereotyping is sort of the derogatory label for when the stereotyping is bad. Oh, it represents diversity in a negative light. Oh, that's interesting, Sonia. I don't know if you have anything else you wanna share on that, that's interesting. Maybe because of the image of the woman in the middle. You might also have noticed, yeah, using the word buck, you notice the image on the left, it says young buck, young man. My read of that face is that it's not male. Caitlin says very rigid categorization space and things like short hair. Yeah, exactly right, 100%. I don't wanna know what image network categorize me as. I must confess. So these seem kind of problematic categories for these images and they're very negative. Well, I don't know, I should take that back. It's not that they're negative necessarily. I think sick persons, what is it? Sick person, diseased person, sufferer. That seems sort of negative. The image on the left I think is misgendering the person. Again, we don't know. Maybe they, who knows, right? What their own identity is. But it's an assumption, shall we say, that this person is male or female. And then the middle, we've got smasher, stunner, knockout, beauty, attractive, seductive looking woman. Okay, so let's look at a few more. Uh-oh. That's somebody who posted there after they fed their image into image roulette. They then tweeted about it and took a screen capture. It's interesting to me that you see these categories of rape suspect or first offender. And again, remember, these labels, the categories themselves are coming from a categorization schema, if you will, from WordNet that are then being used by humans who are labeling the images for ImageNet. And so, hmm. So it's not just white people whose faces ImageNet might classify into problematic categories. Let's take a look at how ImageNet's algorithm categorizes dark or black-skinned faces. Her arrow, which, according to the definition, the definition comes from WordNet. A member of a pastoral Bantu people living in Namibia, Botswana, or Angola. Okay, it predicts the persons from an ethnic group in Southern Africa, which I think is sort of actually kind of odd. It's not accurate. But how about another image? Does it get worse than that? Okay, so we've got two images here. More classifications of being black. So maybe ImageNet isn't so bad for dark-skinned folks. So what Crawford and others pointed out is they were experimenting and exploring ImageNet through ImageRulet is that people with white skin received a much wider variety of potential categories or attributes assigned to them than people with dark skin or black faces. They tended to be categorized very narrowly and simply by the color of their skin. Now, ImageNet, the creators of this, they realized in 2019 that there was a problem because ImageRulet made a lot of publicity for them. And so they spent a bunch of time trying to fix the problem. They went through and purged several problematic categories in that person bucket. So here's an example from the paper that I shared the image for of different labels. And again, just looking at these labels in WordNet, drug addict, junkie, narcissist. I mean, what does a narcissist actually look like? I find all of this Hormaster, clown, buffoon, okay. Violator, lawbreaker, law fender. Hundreds of thousands of images were classified of people with these categories and ImageNet's like, okay, maybe we should fix this. So they got rid of the categories and then recategorized the images in their dataset. They also worked harder to offset. I'm sorry, that's not what I meant to say. They also worked to sort of balance the categories better around age, gender, and skin color by feeding into the algorithm a broader representation of say, genders, age, and skin color so that there was more accuracy in the classification of the images of people. And I see some comments. I'm just gonna go back and take a quick look here. So Sunny said, this would make sense since people of color in terms of the category schemas are not thoroughly explored or defined as extensively. Yeah, I think that's a good point, Sonia. And then Caitlin, you wrote, I agree it highlights the lack of representation that black people have, absolutely right, 100%. Good. So I'm gonna sum up here what the problem is as I see it and not just me, and I'm gonna give some attribution here in a second. So first of all, wordnet, which again, is the categorization scheme that was used to apply to the images. It not only includes the names of objects and the relationship, it's also providing evaluations or judgments on those objects because actually that's how meaning works. And ultimately what wordnet was trying to do is to build associations between words to help develop meaning. So if you've used chat GPT and if you haven't, I highly recommend that you do. If you start playing with chat GPT, you will see how powerful it is at creating sentences and paragraphs that make perfect sense because of the efforts by people who'd built wordnet going forward to begin to see how words in relationship to each other create meaning. So there's power in this and that's what I really want you to sort of think about how naming is powerful. Oops, I need to fix something once I, there we go. All right, so wordnet's people category includes associations of people that again, have these value assumptions like you all highlighted quite nicely when we were looking at those images. So bisexual is categorized under a sensualist category but that makes assumptions a whole host of assumptions by the creators of wordnet about what it means to be bisexual that there's something extra sexual about somebody who's bisexual that somehow isn't the case if you're heterosexual or queer. And that again, that's really problematic. ImageNet reproduces and amplifies these problematic associations and labels by serving as the basis then of these algorithms that help for image extraction. So and then ImageNet and their datasets are building algorithms that also have some problems with categorizing for example, dark skinned people solely by their skin color. And the concept for that is digital epidermalization which is this idea that if you are dark skinned you are more likely to be labeled or identified simply and solely by the color of your skin rather than the other possible attributes. So remember when we looked at the woman at the beginning, the black woman you offered other categorizations like professional or independent and those in these classification systems are often not there for people of color. ImageNet's dataset also it's worth mentioning is a Western dataset. So the images are Western, of course these labels the categories reflect Western culture and the ways we think about the classification and the attributes of objects and that matters. And then finally both WordNet and ImageNet creators at the end of the day, Dr. Lee and the creators of WordNet were more concerned about the engineering challenge than they were about the social implications of what they were creating. Oops, that is not what I meant to show you, hold on a second. So the implication of ImageNet and WordNet still lives on. So if I were to do a Google image search for woman, oops, now I've broken my thing, there we go. Here's what I get. What do you think? So this is woman. How does this look to you? What do you think? Yeah, it's more diverse. Really look at some of those images. What do you see? So NATO says, I'm shocked because I see four black women and age diversity. Yes, agreed. We're gonna talk a little bit about beautiful in a second. That's coming. More representation, different cultures. So Sony notes. I believe that I also see transgender women in this set of images. And I can tell you that Google has actually spent a lot of time with some of these main categories like woman to basically kind of put their thumb on the scale to make sure that they're bringing a more diverse representation to these main categories when you search for them. But as Lily noted, right? So up here at the top, you can see these subcategories, chest, I'm not clicking that one. Beautiful beach girl, pregnant bed, bra, perfect. For fun, I drilled down from woman to beautiful and then also to face. And if I do that, here's what I see, right? Black and brown women are essentially erased from the category as are transgendered and older women. They're not here in a beautiful face, a woman's beautiful face category. So they do essentially look the same. Maybe they're two fair-skinned people of color. I think that's fair. I couldn't quite tell, to be honest. But if they are, oh yeah, maybe there are two there. They are very light-skinned, right? There might even be an AI person in this. That's up on the upper right, I think might be. So yeah, and they're definitely all young. I am not in this, right? So I don't get to count in this category when you drill down. So Google has power in how it categorizes people and who gets to be included in the category of beautiful women's faces. And some of us in that classification scheme, we don't belong. So if you found this interesting, I encourage you to check out these readings. In particular, I wanna call out Dr. Kate Crawford. She's the one who I mentioned around the image roulette activism project. She writes about image roulette as well as the case of ImageNet in her book of Atlas of AI. I also recommend the other books, Algorithms of Oppression by Safia Noble highlights how black people and black women in particular are erased in search engine results. So I have enjoyed the conversation with you. I appreciate your engagement with me in this. I'm gonna leave you with this question. Will it get better? I think that depends on you guys.