SciShow Tangents - Machine Learning

Episode Date: April 30, 2024

We do a lot of learning on Tangents, so naturally, we've got questions about what learning looks like...for machines! AI and machine learning science communicator Jordan Harrod (https://www.youtube.com/@jordanharrod) lends her expertise to this episode as we definitely for sure learn all there is to know and will ever need to be known about machine learning, algorithms, and AI. Definitely we got all of it, everyone.SciShow Tangents is on YouTube! Go to www.youtube.com/scishowtangents to check out this episode with the added bonus of seeing our faces! Head to www.patreon.com/SciShowTangents to find out how you can help support SciShow Tangents, and see all the cool perks you’ll get in return, like bonus episodes and a monthly newsletter! A big thank you to Patreon subscribers Garth Riley and Glenn Trewitt for helping to make the show possible!And go to https://store.dftba.com/collections/scishow-tangents to buy some great Tangents merch!Follow us on Twitter @SciShowTangents, where we’ll tweet out topics for upcoming episodes and you can ask the science couch questions! [Truth or Fail Express]Racoon-inspired algorithmhttps://ieeexplore.ieee.org/abstract/document/8552661Dolphin-inspired algorithmPelican-inspired algorithm[Trivia Question]Using machine learning to measure plant vibration and sounds when they experience stress https://www.cell.com/cell/fulltext/S0092-8674(23)00262-3https://www.smithsonianmag.com/smart-news/plants-make-noises-when-stressed-study-finds-180981920/[Fact Off]Lenna (an image of a Playboy model) used as the standard for digital image processinghttps://www.404media.co/lena-test-image-ieee-policy/Tricking speech recognition systems with “neural voice camouflage”https://www.engineering.columbia.edu/news/block-smartphone-microphone-speech-recognition-spyinghttps://web.stanford.edu/~jurafsky/slp3/16.pdfhttps://arxiv.org/pdf/2112.07076.pdfhttps://www.vox.com/2018/7/20/17594074/phone-spying-wiretap-microphone-smartphone-northeastern-dave-choffnes-christo-wilson-kara-swisherhttps://gizmodo.com/these-academics-spent-the-last-year-testing-whether-you-1826961188[Ask the Science Couch]AI image generator mistakes (hands, feet, teeth) and how they work (diffusion models)https://www.sciencefocus.com/future-technology/why-ai-generated-hands-are-the-stuff-of-nightmares-explained-by-a-scientisthttps://www.buzzfeednews.com/article/pranavdixit/ai-generated-art-hands-fingers-messed-uphttps://hai.stanford.edu/news/what-dall-e-reveals-about-human-creativityhttps://proceedings.mlr.press/v37/sohl-dickstein15.pdf[Butt One More Thing]Training machine learning models (convolutional neural networks) to classify poop imageshttps://arxiv.org/pdf/1903.10578.pdfhttps://journals.lww.com/ajg/abstract/2022/07000/a_smartphone_application_using_artificial.24.aspxhttps://pubmed.ncbi.nlm.nih.gov/33275399/Anal Recognition Paperhttps://www.nature.com/articles/s41551-020-0534-9

Hello and welcome to SciShow Tangents, it's the lightly competitive science knowledge showcase. I'm your host Hank Green and joining me this week as always is science expert and Forbes 30 Under 30 Education Luminary, Sarri Riley! Hello! I feel the most anxious about my epithets when we have an actual expert on the podcast. And our resident everyman who doesn't have to feel anxious at all, Sam Schultz.

Yeah, I still do all the time. And today, we do have a very special guest. It's a PhD candidate researching brain machine interfaces and machine learning for medicine and content creator on YouTube and Nebula who professionally turns her hobbies into work. It's Jordan Herod. I know a thing or two about that. Hello. Hello. I am also probably professionally anxious. So I feel like I'm in a room with friends. I don't think you can be a YouTuber without being anxious, unless you are a bad person. That's the direction it goes.

Interesting, okay. I have a question. I guess I won't spoil it, even though it's in the title, but I have a question. So, when you're using like a chat GPT or something like it, sometimes it apparently helps if you like bribe it or threaten it or explain a certain situation to it that like it's a star ship captain

and has to correctly plot the course to the center of the nebula. So it must do the math questions correctly. If you were a generative AI language model, what would people have to do to you to threaten you to get you to output correct information? They'd have to tell me I get to lay down after I did a good job.

I feel like they would have to tell me that I can't take my ADHD meds the next day if I don't. Oh, yeah. That'll really kick the hyper focus into gear. Yeah, but then if you didn't do it, they'd ruin their model the next day. Oh, absolutely. They take that risk.

I think they would just have to say that they'd be disappointed in me if I didn't do it, which is like the fuel that drives me so much in everything that I do. I think they could say, we'll be totally fine. You'll be disappointed in yourself. Oh my God. You'll be disappointed in yourself. It actually doesn't matter to us one way or the other. I'll be at 4am trying to find your answer to the math question. And then you'll say it and then it'll be like, I forgot I even asked you that.

Thanks, I guess. They just leave like one comment on the YouTube channel that isn't like explicitly negative, but it's like generally questioning the intellect and implying that I am wrong. I will be down that rabbit hole for the rest of the day. Oh, some soft criticism is the worst where it's like, I just don't really think that they that like the thing that they're doing is for me. And I'm like, but I meant for it to be

very good and okay job explaining batteries, but she missed some things. And I'm like looking up courses on electrical engineering. How? How? Re-enrolling in college. Yeah. Why did I do this? Oh, I'm too emotionally stable.

I just want a hot chicken sandwich. That's I would do. I would answer any question for a Nashville hot chicken sandwich. Yes. That's all I want. I would answer any question to be emotionally stable. Yeah. A mood. Sandwich. Yes. That's all I want. I would answer any question to be emotionally stable. Yeah.

A mood. Oh, well, I can't offer you that. But we're we're working on it as a society. That's the goal of this podcast. Someday we'll get there. We're all going to be better people. And then that'll be our last episode. We don't need to impress anybody anymore.

Yeah. And then we just walk through the gate in the woods and cease to exist. That'll also be me with you too, because I assume that the AI situation will just be fixed. Mm-hmm. That sounds exciting to get to the point where the AI situation is fixed. Fingers crossed. I feel like that's a while yet into the future. Every week here on SciShow Tangents, we get together to try to one-up, amaze, and delight each other with science facts while also trying to stay on topic.

Our panelists are playing for glory, but also for Hank bucks, which I'll be awarding to them as we play. And at the end of the episode, one of them will be crowned the winner. But as always, we must introduce this week's topic with the traditional science poem, this week from Jordan. It's hard to know the ways you're blinded by the cause you hope to fix. If one can barely comprehend, the wealth of knowledge stored as bits. Shaping a system line by line and stumbling through a field of minds, predictions come not always true,

but data's costly, we must make do. Released into the light of day, come listen to what it has to say. Hope that it's making the right choices and amplifies the unheard voices." And while this poem isn't friendly to companies that do offend me, it's true that they're not moral actors. There's more than what they manufacture. It's not a person and not your friend, but it probably won't cause the world to end. It can be helpful, don't forget. Careful design can keep your values met. This poem was not made with AI, but with two hours and a glass of wine and I couldn't find a place to say the only winning move is not to play

Oh, I'm gonna lose then I mean, I'm hoping that every AI movie does not Come to fruition in real life because we're we're we're gonna be in a real bad place in that case Yeah to fruition in real life because we're gonna be in a real bad place in that case. Yeah. If anything I've learned from science, reading science fiction and being alive for 43 years is that it always ends up much weirder but also less interesting. That's totally true. We're not gonna get to go anywhere fun or have any adventures. We'll just watch strange things happen from afar. Instead of Mars, we made Facebook. Yep.

Good job, everyone. So this week's topic is machine learning, but before we dive in, we're going to take a short break and then we'll be back to define the topic. All right, Sari, are you going to try and define machine learning in front of Jordan? Because that sounds terrifying. I'll do my best. But Jordan, you can jump in at any time to correct your knowledge. So artificial intelligence is like a broader umbrella. And there is artificial intelligence

that involves machine learning. And there's some that doesn't involve machine learning. And machine learning, in my understanding, as not the science expert of the episode, is when you develop algorithms, usually they're statistics-based or frequently they're statistics-based on a machine, which is where the machine comes develop algorithms, usually they're statistics based or frequently they're statistics based on a machine, just where the machine comes in, like a computer, that can learn from data

and then use that information to process new data on their own without explicit instructions necessarily. So you might be able to code something or code a program that says, if this is true, do X, if not, do Y. If you find a word that starts with A, add one to this counter. If you find a word that starts with B, don't add one to this counter. That is not machine learning. But if you train a machine on a set of words so that it can make predictions about what word goes next in a sentence. That is a form of machine learning. And then there are different ways to teach these algorithms.

So you can do it in a supervised way where you provide a labeled data set. So that is if you provide a bunch of images and you tell a computer that those are all cats, and then it learns. You give it both cats and not cats. Yeah, give it, I guess all cats and then it learns. You give it both cats and not cats. Yeah, give it, I guess cats and not cats. It's like, here's a bunch of cats, here's a bunch of not cats, and these are cats and these aren't cats.

Then you're like, is this a cat? And it's like, yeah, or no. The thing that bugs me about all of this is we have no idea how any of these things actually work. We know how we build them, but we don't know how they're deciding what to do or say. Or we don't... Like nobody knows that.

It doesn't know what a cat is. It just like knows that this is a cat, which are two, they sound like identical sentences, but aren't. Yeah, a lot of it is finding patterns in data. So there's some sort of representation of a cat and there's some sort of representation of not a cat. And that's the thing that it's learning to discriminate between.

But like what that is, we don't necessarily know and how you like translate that into English words is particularly challenging as models get bigger because you can, you know, theoretically explain like you put in, I don't know, the square footage of a house and then the price of the house and you find something to model that and that can be machine learning. But that often creates some sort of formula that you can interpret in some way versus chat GPT where you could open up that box

and it would just be staring into the abyss. I'm gonna ask a stupid question because that's my job. Is that distinct from how people understand stuff? Could we say that we understand how we understand stuff? So I think there are parallels between how we learn and how models learn. I could go into a whole deep dive on like why neural networks are called neural networks and how they're based off neural architectures and the

neuroscience side of my brain will then come in and explain why brains are real complicated and we do not know that much about them and so like information storage parallels there don't necessarily translate that well because it's not really a one in, one out situation. It's a lot of things in and a lot of things out in very complicated ways. So it's not unrelated to how we learn. We do learn faster.

So I guess when people talk about like zero shot models, which is kind of a common term in the ML space, what we're talking about is having a model do something based on input that it has not seen and it being able to kind of infer how to process that information even though it hasn't seen the exact representation. And so it takes a lot to get to a point where a model can do that. That's something that as we've seen like the exponential curve of And so it takes a lot to get to a point where a model can do that. That's something that, as we've seen, like the exponential curve of AI go up over the

last five to six years, has been something that's easier to achieve. But it took like that first three and a half years to get there. And now we're like going all the way up. And babies can do that in like six months or less. So humans pick it up faster. But the species of computer, right? Like the species is learning faster than the human species learned.

Is that true? Like Linux versus Mac or like? No, no, I think what Sam was saying is like, it took several billion years to get from a single celled organism to a human. Correct. Yes.

Thank you, Hank. Yeah, going from the first like vacuum tube transistors to now is quite quick, which is sometimes terrifying. Like, part of me wants to feel very comforted that like, oh, well, this is still not as good as me at things. And I assume that that like last step is much bigger than people think it is. Yeah, but I don't know. I don't know what they're working on next. Like, I wouldn't have thought that if you put a bunch of sentences

into a machine learning model, that you could then have a thing that would convincingly be able to answer the question, why is a vegan riding an electric bike so punchable? Oh, it does know humor and sarcasm these days, which, um, yeah, it's going to steal all of our good jokes. It's going to be the thing that takes us out. I thought it was good that artificial intelligence, like the heart, the thing that would be worst at was jokes because I was raised watching Star Trek. But it turns out that's no problem. Whereas 53 plus four is really tricky.

Why is that? Is that true? Yeah, they're bad at math. What the hell? Because it's like, they're not programmed to do math. They're like language models. They're programmed to predict the next word in the sentence. And like, people don't walk around saying 53 plus four equals 57 all the time. Took me a second. And like people don't walk around saying 53 plus four equals

57 all the time took me a second That would be a much less useful data to have than the data from Star Trek if data was just like a stand-up comedian Yeah, you crashed your spaceship all the time. Yeah, but like bad at it. You only told jokes that other people have told before Yeah, oh my gosh. Well, I'm glad that we've gotten to the bottom of machine learning, everybody. I feel like we've really sort of got it all down. No stone left unturned. Everybody understands it as well as is going to be necessary for the future of being a human on a planet where things are constantly changing extremely quickly.

Yeah, I think we're good. Why am I doing this PhD? We're going to move on to the quiz portion of our show. We're going to be playing a little game. It's called Truth or Fail Express. Here on SciShow Tangents, we often bring up technologies that are inspired by nature. Scientists have figured out all sorts of cool materials

and machines and other things that cheat off of biology so we can figure out how to approach different problems. But you know who also has problems to solve? Animals do. And people working in machine learning have developed all sorts of methods inspired by the way animals solve the various problems in their lives, including ants, killer whales, and gray wolves. So today we're gonna be playing Truth or Fail Express. But with a little twist in honor of our theme, I will present to you an algorithm inspired by an animal behavior that was stitched together entirely from lines written in an academic paper. Or it might be just from chat GPT. So you have to tell me if it's a real paper or chat GPT talking about weird animal-based algorithms.

Like for example, the raccoon optimization algorithm. Do you want me to tell you about it? Please. Yes, please do. Raccoons are known as intelligent and curious creatures. These qualities, combined with their dexterous paws, make raccoons extremely successful in searching for food. Moreover, zoologists believe that raccoons extremely successful in searching for food. Moreover, zoologists

believe that raccoons have an excellent memory. The process of finding the global optimum for a predefined fitness function in the raccoon optimization algorithm is inspired by the natural food-searching habits of raccoons. Food, which represents different possible solutions of the fitness function, is spread throughout the raccoon's living environment. The algorithm makes use of two different search zones in each iteration. In addition, the raccoon can remember best previously visited locations. Then, if it does not manage to find better solutions in future iterations, it can revert

back to the best location. Is this from an academic paper about machine learning or from chat GPT pretending to be an academic paper about machine learning? That to me feels like it was written by somebody who knew what they were talking about. And I just simply don't understand what they were talking about. That is how I felt reading it. I'm going to go academic paper. The moreover was where I started zoning out and it was also my like, this feels like a

chatbot maybe. So I'm gonna go with chat GPT. It is pretty wordy for an academic paper in CS. That's all the writing that that researcher did that year. I guess I'm gonna stick. I'll stick with paper. I just feel like a chatbot maybe would try to spice it up a little bit part way through it.

It would do better? Yeah, it would do better. It would write something wrong but more interesting. Well this text is from a paper published in 2018 titled Raccoon Optimization Algorithm. The researchers took inspiration from the fact that raccoons are known to be very good at solving problems, namely the problem of finding food. And the idea is to help the raccoon explore the area around it, starting by defining a reachable zone where it can look for potential solutions. But there's also a visible zone where raccoons can keep an eye out for other solutions.

So that's a real thing, and I don't know why I we wouldn't because they seem Have you guys ever had the thought if raccoons had gotten sentient first? They probably would have blown us the world up faster than us Yeah, cuz it gives me some comfort to know that we probably weren't the worst thing Like I feel like raccoons probably would have done a worse job than us There's gotta be an anxious raccoon out there who would feel bad There's plenty of mischief makers, but there's always probably a guy saying, what if we don't steal the trash? What if we're just like, are a little bit nice? What if we tidy up a little bit?

Guys, we're going to get in trouble. Yeah. You got to hope that that raccoon becomes the president of the raccoons and then everything's okay. But odds are, odds are not gonna happen. As we'd say, if they're doing worse than us, then not holding my breath on that, I don't think. I like the idea of just like a raccoon president standing on top of the trash being like, I'm the craziest one.

You picked the craziest one. Just punching all the buttons. It's like we do, you know? Yeah. All right. The next algorithm is dolphin-inspired. Dolphins possess exceptional cognitive abilities, including highly developed echolocation skills and complex social behaviors.

These capabilities enable them to navigate challenging environments, communicate effectively, and cooperate with peers to achieve common goals. The dolphin-inspired learning algorithm, DILA, incorporates an echo-based sensing mechanism to capture information from the environment, simulating the echolocation abilities of dolphins. This mechanism enables the algorithm to extract relevant features and navigate complex data landscapes effectively. Inspired by the social behaviors of dolphins, Dilla employs collaborative optimization strategies to facilitate information exchange and collective decision-making among multiple agents. By leveraging the collective intelligence of the algorithmic ensemble, Dilla enhances

robustness, scalability, and adaptability. Chat GPT or real research paper? This is rough. There was like the scandal where a biology research paper got published with figures that were generated. Yeah. Oh my God. It was barely a scandal.

Those figures were amazing. I want to get that Tattooed on me. It was like how big can you we make the mouse penis? Yeah I'm searching mouse penis right now I'm searching mouse penis right now. It's a lot shit That's my metric for AI papers Yeah, if there was a visual cue of a dolphin with a giant penis. I would be like that was AI obviously Unfortunately, we have just the text read to us by Hank to go off of, which is so much less of a visual indicator. Would a visual component help with this that much?

I don't know. I mean, if it looked like this mouse. See, these are the kind of buzzwords that you get in front of a boardroom and you start saying, moreover, you start saying whatever they say. You start saying robustness, scalability, and adaptability. And at this point, all of the people with the suits on are going, yes, and we're writing checks to you, sir. And sir, in this case, is a robot because a robot wrote this so that people would be

excited about it. My strategy is I'm just going to get us to chat GPT every time and hope that one will be right, maybe. It's got to be, I assume. See the problem is that wait, what was it called Dilla? D-I-L-A Dilla. That feels like a bad acronym. I'm gonna go chat GPT I wouldn't I believe that this could be a paper that someone wrote on like archive But a scientist would be having more fun with that acronym

I think you know a real person now it would be some I feel like the thing in CS is finding the Quippy title. So like once attention is all you need was the big transformer paper. Everyone had to riff off of that. So it wouldn't be an acronym. It would just be something like snarky. Well here's the prompt Deboki used to create this paper. It said, make up a machine learning algorithm inspired by dolphins in the style of an academic paper. That's what it hit us with. It said, make up a machine learning algorithm inspired by dolphins in the style of an academic

paper. That's what it hit us with. That was ChatGVG. All right. Last one. Pelicans after identifying the location of the prey, dive to their prey. Then they spread their wings on the surface of the water to force the fish to go to shallow water.

The fundamental inspiration of the proposed pelican optimization algorithm is the strategy and behavior of pelicans during hunting. In the first phase, the pelicans identify the location of their prey. In the second phase, after the pelicans reach the surface of the water, they spread their wings on the surface of the water to move the fish upwards. This strategy leads to more fish in the attacked area to be caught by the pelicans and causes the proposed Pelican Optimization algorithm to converge to better points in the hunting area. Is this from a paper about machine learning or from a chat GPT pretending to be a paper

about machine learning? What are they? What is the algorithm doing? What is the fish? Are the fish in the computer somewhere? The fish are the captains. What are the fish? The problem is that these are such long descriptions.

It's like no one writes this long. It doesn't look as long on the page. But when I'm saying it out loud, it takes forever. Just go forever. I feel like this I can picture what the pelicans are doing. So it's like, that's logical. But you're also a pelican guy. So it makes sense that the book you would be like, it is.

It almost feels like a red herring to have that. And yeah, exactly. But they don't say what the fish are to have that. Yeah, exactly. But they don't say what the fish are. And that's a real problem. Or what the pelicans are. I'd probably get to it later in the paper. You can picture what the pelicans are doing,

but you can't picture what the machine learning algorithm is doing. And so I think it's chat-shitty-titty again. It's pushing the data upwards. Yeah, I was about to say, they they dive and then it pushes the fish up. I'm gonna go paper, you know? Let's... It would not be the weirdest thing that I've read. What I'm remembering sticking point to me is that I don't know...

I don't know that pelicans hunt collaboratively. I feel like I've only seen that... I will tell you that in fact white pelicans do hunt collaboratively. They do. That's clearly that they're talking about the hunting strategy of white pelicans here. But are you telling me this? We're not plunge feeders. Everyone thinks that all pelicans are plunge feeders, but they're not. There's only two plunge feeding pelicans. I think you gave me that piece of information.

So I guess paper. So I'm going to guess Chad GPT. Oh, wow. No, I will just tell you as much about Pelicans as I possibly can. OK, that is a fact you should know about me. This was a real paper. Oh, man. Wow. Jordan, you are in the field. You are nailing your peers work. It's a it's a 2022 paper.

It was titled Pelican Optimization Algorithm, a novel nature inspired algorithm for engineering applications. The algorithm is balancing the way that pelicans explore for food, which is, but I don't know. I don't know what it's doing. No one explains to me in the paper, in the description what the heck it's actually doing

and what the fish are. Yeah, I definitely still don't get what the fish are. The fish are solutions is what Deboki has written. In this case, the fish, it's sweeping solutions into an area. It's like a political cartoon with just way too many labels. The fish are solutions. Yeah, Raccoon and Pelican both sounded like versions of reinforcement learning slash control problems,

which made sense as papers. And then dolphin echolocation came in and that one just like threw me so much that I was like, I don't know. It did seem like it was both like suddenly halfway through the introduction it was like, and also it's about dolphin social structures. Like it's not just echolocation,

it's a completely separate dolphin thing. That'll be the 2025 paper. We will create consciousness by mixing echolocation so that the thing knows about physical space with social structure, which is... And then they'll all sing, So long and thanks for all the fish. And they'll take off into space and be like,

Look, you guys can't do AI anymore. We do not trust you. The moment you start to code again, we got a laser up here ready to shoot you. All right, everybody. Jordan's in the lead with three. Sam's got two, Sarah's got one. Next up, we're gonna take a short break

and then it'll be time for the Fact Off. [♪ Music playing. Welcome back everybody. Now get ready for the Fact Off. Our panelists have brought to me science facts to present to me in an attempt to blow my mind. And after they have presented their facts, I will judge them and award Hank Bucks anyway I see fit. To decide who goes first, though, I have a trivia question. Research has shown that plants vibrate when experiencing drought. And scientists, that's real

Okay, let's keep reading to find out what's going on scientists look to see what kind of sounds that might produce recently researchers set up microphones about four inches away from tomato and tobacco plants and recorded the plants after cutting them or Not giving them water and they found that the stressed out plants made more sound compared to the non-stressed plants. While the sounds were at a frequency too high for humans to hear, the scientists did design a machine learning model to distinguish between the sounds produced by the cut plants and the non-water plants. How accurate was the model at differentiating between the two stressful conditions in a percentage?

I'm still not over the fact that plants vibrate. Don't they don't look like they vibrate to me! I feel like I wouldn't notice that, but I guess it's a little vibration. I would vibrate if I'm stressed because I'm thirsty. Oh yeah, I totally vibrate when I'm stressed. Or cut. How accurate was the model? I feel like the most accurate models are in this 70%, 80%,

maybe, that I've seen. But I also am not in this field. So that would be like high accuracy, in my opinion. Maybe like 62% of the time it was able to differentiate dry from hurt. I'm going to go 75. The answer was 70% so Jordan gets to go first. 70% is also, it's somehow both lower and higher than I'd expect. I don't know. So my fact is when digital image processing research was getting off the ground in the

70s, the standard test image, so the image that was used when people published papers, when they went to conferences in order to standardize everyone's results, was called Lena. And it came from a Playboy magazine centerfold that one of the researchers just happened to have on his desk at work. That research is foundational to pretty much all image and video-related AI results that we see now,

including image generation, video generation, deepfakes, et cetera, and then a not so shocking turn of events. The model received exactly zero compensation other than what she was paid for the Playboy shoot, nor did she consent to have her image used in that way. Although she would go to conferences every few years and be like, yay, I'm happy that you guys are finding this useful. And it wasn't until I think the early 2000s that she was like, okay, can we stop doing this? That's not enough.

I think it was only like last week actually that IEEE, one of the big conferences in the field was like, okay, you can't use this photo anymore. It's slowly been phased out of the ML community as something you are allowed to submit with. I'm looking at the photo now. I do recognize it. I feel like I have seen this picture before. It is, maybe important to say, in an audio medium, just from the shoulder up. So they didn't, they did not do this with an actual full-body naked woman. So there's at least

that, but they did do it with a woman who was naked at the time of the photograph being taken. Correct. That's wild. And so this is like with an image, a standard image they would use and try to reproduce with machine learning and that was like the... So not reproduce. So when people were trying to create

and test out different methods of processing images, so like filtering images, transforming images, it was easier to have one photo that everyone could use to compare results and methods on. Gotcha. And so this was the one photo. So if you're like all those early Photoshop things, you know, increasing the edges and the making it fuzzier

or changing the hues and saturations. Yeah, gotcha. I guess there's probably not any way have any compensation actually occur at this point. I would imagine not like academia doesn't pay people in academia particularly well, let alone everyone else. Like a bunch of money. So I don't know where that money would come from.

Yeah. Adobe. I don't know. Yeah. I was about to say, if Insert X major AI company would like to compensate her for her contribution to the field, then they should be out. Might be just a good press play. Yeah.

Like, it's just like, hey, you want some good PR today? Here's this lady whose image was... or her contribution to the field, then they should be out. Just a good press play. Yeah. Like, hey, want some good PR today? Here's this lady whose image was, but also probably the Playboy company owns the picture, actually. Playboy would go, uh-uh, now wait a minute. Just give people money. This is how I feel. People who do weird stuff, if you're

a company with billions of dollars, just make their day. Just mail some checks out, guys. Make somebody's day. Why not do that? I support this, especially if the check goes to me, but also everyone else. Are there other images that are used now as standard image, like that replaced this, or is it at a point in the field where you don't need this sort of like anchor digital image processing

reference. As the field has evolved and as we've gotten into like image generation, deepfakes and things like that, there's been a little bit less of a need for standard images because like you don't need a standard deepfake image to work with. I think what's happened more so is that we have ImageNet and the really big image datasets and that's the standard thing that everyone uses to build their models and then you can test performance as a way of having the same baseline. I guess I knew that that would be the case.

And also in the same way, all of those images are scraped from the internet without the permission of the people who made the birth in them. So you know, 50 years later. We're still at it. The proud tradition lives on. All right, wild, interesting. Sarah, what do you got? One of the ways that machine learning is being applied

right now is called automatic speech recognition, which is basically one of the driving technologies behind any voice assistants, like Google's assistant or Alexa or my mortal nemesis Siri. And my very basic understanding of what these algorithms do is they take the sound wave form of your speech and then chop it up into very small pieces. And then the algorithms at work try to match

those tiny pieces to sounds that they were trained on. And by putting those sounds back together and incorporating other information from training data about different words and how words become sentences, that makes sense. The computer does its best to try and guess out what you're saying. And this is a hard thing because people mumble

or have different accents or misspeak or whatnot. So the same sequence of phonemes doesn't necessarily translate to the same sentence. But as these voice technologies are in more places, people are worried about AI listening to them or surveilling them. And I feel like this is a huge paranoia that I see pop up every once in a while, is that if you talk about with your friends how much you want chocolate milk, then you'll start getting milk ads across your devices or something like that.

And as far as I can tell, a lot of those targeted ads are more about what you've been searching, like if you're searching straws for chocolate milk, or if you're sharing a Wi-Fi network with someone who's searching those things, or if you were physically at a chocolate cow dairy farm or something like that, that can provide that location data rather than listening. So data privacy is, it's kind of tangential. Data privacy is a concern, but your phone probably isn't recording and transmitting everything you are saying all the time.

A rabbit hole I went down as I was trying to find this fact. But the fact is, even still, as an experiment in audio counter surveillance, there was a paper from Columbia University researchers that was presented at a conference in 2022 that showcases a machine learning model that tries to disrupt speech recognition algorithms in real time as someone is talking. So it basically uses similar principles to take your voice as an input. It does a series of calculations on it. And then instead of interpreting the speech, it instructs computer speakers to emit specific noise. The press release says it's whisper quiet, but there weren't audio

samples so I couldn't hear it. And the noises specifically exploit the weaknesses of existing speech recognition algorithms. So they change the waveforms in the room just enough either by adding in interference or adding like blanket white noise or other things like that just enough that those chopped up sounds don't get interpreted as sensible phonemes or words or sentences anymore even though they're fine by human ears the computer can't interpret them. And they tested this model across various speech recognition programs and databases I think kind of like Jordan was saying,

my understanding is that these are open source and like regular databases that people train their models on like Deep Speech or Libra Speech or Wave 2 VEC 2. And if I'm interpreting the figure in the paper correctly, this counter surveillance thing blocked anywhere from 28% to 87% of recognizable speech depending on the system. So it's not perfect.

It's more for a hypothetical spine situation than a real one. But I think it's interesting to talk about the flaws in all these technologies as much as the breakthroughs. What if like when I'm doing my spy craft, I just talk, I'm just like, hello, it's Hank. I think the computer would just put a big old question mark. Yeah. That's how it works. I think whoever's listening would also put a big old question mark.

So if you just like had this on in your house, all it would like at the beginning of your fact, basically you said your phone's not actually listening to you in order to advertise to you. So to have it on your house, all you'd be doing is making it so you couldn't say, Hey, Siri, and it wouldn't wake up. Right? Yeah, I think effectively, yes. But it was weird because even though there's this body of research that's

saying that your phone isn't spying on you at the beginning of this paper, they were like in situations where you might be monitored by technologies around you. And so I don't know if this is, I fell down that rabbit hole when researching whether that was like a legit thing. I would also like for YouTube to add that to the background of every YouTube video in which someone says, hey, whatever. So it stops setting off my various devices at home.

All right. I just love the weird esoteric history bits, like things that are like so weirdly important and will inevitably be forgotten, but are also just shine a light on the bizarre culture that that results in many of the in all of the structures around us. So I must announce Jordan as the winner of the episode. She was also way ahead, very powerful lead. It's funny, cause there was an article about this that literally came out last week. And so for the last week, I've been most days going through your tweet likes and your replies

to make sure that you didn't see it. Oh no, no, what did I like? I like some real stinkers I bet. Hank is the one being surveilled all along. I was just reading for discussion of that article. Okay. I cannot comment on anything else that I saw. And now it's time for Ask the Science Couch, where we ask a listener question to our couch of finely honed scientific minds. At Ponyoty on YouTube asked, why does it get human extremities so very wrong? By that I assume it means AI that draws pictures. Can I guess? Because I just have a guess. Which is that it's looking at what object is coming next, and the next thing that comes after a finger is usually a finger.

And so it's just like, I guess, because look at all my fingers, there's a bunch of them. And so to know that there's four is a very different thing than being like, what comes after a finger? A finger. What comes after that finger? A finger. And then maybe on the fifth or sixth finger it draws, it's like, oh, another finger doesn't come after that one. But after the third or fourth, it's like, I don't know.

I don't know where the other finger comes next. It's like, that's too many. Can it ever take a step back and look at what it did and be like, Oh, I did two, my fingers or when it does it's six finger as it is it. Yeah. Does it not ever look at it and say, Oh gosh, too many fingers. It's starting to draw the finger and it's like, Oh shit, I'm drawing an extra finger. I got to put the delay and I guess, yeah.

How you feel as an artist, Sam, when you're drawing a guy, no, I can erase stuff. I'm a human man. That's the power I have over AI. And that's how we're going to win against the AI. Some models can do that. It's ultimately down to how the model is built. So there's also I think it was deep mind made language model where it had like inherent fact checking systems where it would like create a response and then it would be like, let me go like pull a bunch of information from repeatable resources and see if what I say matches what these things say. And then if it doesn't, let me go back and adjust what I said. So there are mechanisms to do that when you're

using like, I don't know, Dolly or like Mid Journey or whatever. Most of those don't really run on that. So that's why most of what you see is like weird number of fingers or like there's, you know, three fingers, but then one's coming out like this way. Confidently wrong. Perpendicular to the rest of your fingers. Which I guess depending on what data sets they use. I was born with, I guess, five fingers and a thumb, so six on each hand, because one was coming out of each of my pinkies.

So I would imagine that there isn't like a ton of that kind of data in this model, but like maybe. I did make the mistake of typing in weird AI hands into Google and it is it is nightmare fuel. These fingers have fingers. I don't like that. In all likelihood, it's it's an edge case issue in that there are lots and lots and lots of photos of people's faces and people's bodies in a bunch of different positions and whatnot.

But the the number of different configurations of like hand positions and you know, you may go a lot away obviously have five fingers, but like one might be behind like out of frame or something like that makes it more challenging in the same way that like ears used to be the issue and hair used to be the issue representing that if you don't have enough data, it ends up being really, really wonky in the same way that language models hallucinate things when they don't have good representations of them. The one thing that I do have to add, I guess, is what... And Jordan, you can correct me if I'm wrong because I'm coming in a little bit new here. I think the model that these sort of image generators

are trained on are diffusion models, which were introduced from what I found in 2015 in a paper called Deep Unsupervised Learning Using Non-Equilibrium Thermodynamics, which is a fancy way of saying that they took an idea from physics, so this idea of equilibrium, and that if you have, I don't know, I feel like a pretty classic chemistry problem where if you have two containers filled with particles and then eventually they will like drift over and form equilibrium.

There are a lot of systems that don't form equilibriums, but things drift back and forth. And so you train these models by giving them images and then having them destroy the data in it, which is like an iterative diffusion process or what they call it. So they turn a picture of a horse into a bunch of uncomprehensible pixels. And then the model learns through destroying it how to predict how to create an image, like how to reverse that process and go backwards and generate something from nothing, which is how you get these predictive of, okay, there's a finger, is there another finger? What is the shape of a hand? What is the shape of a letter?

And so that's where the guessing is. That's where the prediction and statistical modeling of these image generations come from and why you can't necessarily, unless there is that fact-checking layer, say this AI knows what a hand is or knows what the letter A looks like or knows what a pipe looks like because it's all just guessing what is the statistically most likely pixel to appear next to this pixel in a certain class of images. And I mean, that's also the issue that you run into with language models doing math.

It's not about the math. It's about what is statistically most likely to come after this sequence of characters. And we as humans are really good at noticing things about other humans. We look at ourselves and each other all the time. And so it's easy to see like that hand is weird. That looks like an alien thing.

Whereas there could be a wrong number of leaves on a tree in the background of a picture and you like wouldn't notice necessarily. Like there's like any botanist probably looks at any AI generated picture of any tree and it's like, oh my God, what were they thinking? Yeah, that flower has an incorrect number of petals. That's biologically impossible.

But because we're all humans, it's really easy to be like, oh, that foot messed up. I've seen a foot and it doesn't go at a right angle necessarily in many cases. If you want to ask the Science Couch, follow us on Twitter at SciShowTangents or check out our YouTube community tab where we will send out topics for upcoming episodes every week or you can join the SciShow Tangents Patreon and ask us on our Discord. Thank you to the Space A on Discord, at Ging-Ging-Ging-Ging on Twitter and everybody else who asked us

your questions for this episode. I did my best. That's what it says. That's right. Jordan, thank you so much for coming on the Sideshow Tangents and sharing your knowledge and understanding with us. I wish we could go deeper. There's so much to know and so much to be excited about and scared by. If I want to see more of what you're up to, where would I go? You can find me on YouTube. You can find me on Nebula. You can Google my name to see what else I'm doing because I'm also working on my PhD and that is taking up 80% of my brain space right now.

Imagine that. And your name is Jordan Herrod. Herrod has two R's. Yes, like the store, which is a reference that only really works when I'm in London. But yeah, I'm like, is it the store? Should I know what the store everybody head on over and subscribe to Jordan on YouTube. And thanks for being here for us. If you like this show and you want to help us out super easy to do that. First, you can go to patreon.com slash size

show tangents to become a patron get access to our discord shout out to patron less acre for their support. There's also lots of bonus episodes and weird stuff, minions, commentaries. Second, you can leave us a review wherever you listen. That helps us know what you like about the show and also other people see them and think, I maybe will watch this show. Finally, if you want to show your love for SciShow Tangents, just tell people about us. Thank you for joining us. I've been Hank Green. I've been Sari Reilly. I've been Sam Schultz. I've been Jordan Harret. Sideshow Tangents is created by all of us and produced by Jess Stempert. Our associate producer is Eve Schmidt. Our editor is Seth Glicksman. Our social media organizer is Julia Buzz-Bazio.

Our editorial assistant is Dibbupetra Krabarti. Our sound design is by Joseph Tuna-Medish. Our executive producers are Nicole Sweeney and me, Hank Green. And of course, we couldn't make any of this without our patrons on Patreon. Thank you, and remember, the mind is not a vessel to be filled, but a fire to be lighted. But one more thing. We talked about some machine learning algorithms are used to sort and analyze different images like dogs, stoplights, or even pictures of human poop. Gastroenterologists already have ways to classify their patients' poop, like the Bristol stool scale or the Brussels infant and toddler stool scale,

which look at things like stool consistency or how fragmented the chunks are. So the goal of these machine learning models, which may eventually become smartphone apps, is to give doctors an extra tool and help patients self-report their gut health more consistently. There was also a 2020 Stanford paper that looked at a similar system embedded in a bidet and it used anal fingerprint recognition to tell which person was sitting on the toilet

to associate that person's data with their selves. That was not in the script. That was just Jordan. It wasn't. I just knew that. Bonus buts. I just knew that. Bonus buts. I did a video about it a while ago.

A wild race. Are people's anuses that unique that you can tell? Apparently. Every one is beautiful in their own way. Specifically in reference to. But who knew that? Who was like, I wonder if the butthole looks different enough. I feel like it was in like the introduction of the paper, because I remember talking about

it in the video being like, who found out? Like who? Yeah. Where's this data set? I have a question.