false
Catalog
The Emerging Use of Artificial Intelligence in Der ...
AI Boot Camp: an Introduction
AI Boot Camp: an Introduction
Back to course
[Please upgrade your browser to play this video content]
Video Transcription
So this is gonna be the boot camp, kind of an introduction for everybody to AI. How many people recognize this scene? All right, some, right? Some from the drill sergeant from Full Metal Jacket. I will be far more kind than this drill sergeant. These are my conflicts of interest and my disclosures. And this image was generated using AI. So today, it's a pretty packed schedule, and we're going to be working together here. I'll be going over some core concepts in AI, and then we'll follow it with Veronica Rodenberg, at Memorial Sloan Kettering, and Albert Chu, my colleague at Stanford. We'll have a question and answer session, and then we're gonna follow it up with Roxana Donashue, Eddie Dovizzi, and Dan Schlesinger. We're gonna have a lightning round after that with some common pain points and some potential solutions using AI. Dr. Ivy Lee, my co-director for this session, is gonna be discussing how to use AI in a community practice. Joe Kovedar will be discussing AI strategy and how to adopt it at a system-wide level. And then Justin Koh will be discussing the future of healthcare. So, as just a basic broad overview, I think when we're thinking about AI, we can kind of think about, just start with, I'll start with my nine-month-old, or with my five-year-old. When they're learning the difference between a dog and a cat, I don't teach them examples, I don't teach them the rules, like, oh, this one has ears that look like this, or it has this, and it would actually be quite difficult to do so. I think that they learn from the world by seeing lots and lots of examples and starting to derive rules from that, and we often do the same thing, where we kind of engage in this unconscious pattern matching to make sense of the world. So, what if we could do the same thing and teach that kind of pattern recognition to a computer? And these are just some melanomas from my clinic. So, the overall field of AI has been around for over 60 years, and we can kind of divide it up into overlapping categories of things like machine learning, where we're kind of trying to teach the algorithms how to think using these examples, natural language processing, where it's learning how to read text and perhaps do translation and other kind of complex tasks, and deep learning, and you can have a lot of overlap between these different categories. So, within machine learning, there's kind of three general, there's more than that, but three general categories that we can kind of think about. There's supervised learning, unsupervised learning, and then finally, reinforcement learning. So, I'll be going over each of these in a little more detail. So, for supervised learning, we basically have, the algorithms are running, and they have an answer for each data set, for each data point. So, it'll show that this is a container ship, this is a Madagascar cat, this is a melanoma, this is an actinic keratosis. This was one of these landmark papers that came out in 2012, with a data set of millions and millions of labeled images, and they trained the algorithm using that. And so, it kind of adjusts over the course of these millions of data points, and minimizes the error across all of those outputs. So, a lot of the AI research that's been done so far has been supervised learning. But you can also do unsupervised learning. So, let's say you don't have all of these labels, and you don't have all of these exact answers. You can actually let the algorithms perhaps cluster the data together, and to see which ones kind of resemble each other. But it's not, again, you don't necessarily have an answer in this setting. And then another example would be something like reinforcement learning. So, this is a branch where it's particularly useful when you have things that you want or don't want, but they don't have a specific answer. So, if you're looking at, let's say, translation, you might have text that is better than another text, but there's not the perfect answer for a translation. If you're teaching the algorithm to play a video game, there might be ways that it'll play that are better than others, but there's not the perfect answer for it, usually. And so, again, this is just another example of the images that you can generate. And I'm showing the prompt above each one of these, to kind of encourage everybody to actually play with these, because it is super fun. I'm a terrible artist, and you can see how neat these images can be, and how relatively easily generated they are. And then we can actually stack that reinforcement learning and put human feedback on it. And so, it can be quite resource-intensive and time-intensive, and it can learn from imitating the way that a human plays the game, or it can show humans different results and different actions, and then it can actually learn from that. So you can, there was a paper a few years ago showing that the same algorithm could learn to play all different Atari games, and could learn and actually end up performing them almost at a superhuman level. And that combination of reinforcement learning with human feedback and something called a transformer has been behind a lot of the recent explosion in generative AI. And so, I think we'll talk a little bit more about this later. But basically, these large language models have exploded since in the last two years, and they've been trained using a large portion of the corpus of text on the internet. And there's an arms race right now, and there was just a recent new salvo in this arms race that came out in the last week from the company Anthropic. So, to just kind of talk a little more, so generative AI is using these large language models and other models to create brand new content. Now, when I say brand new, it's not exactly brand new, right? It's learning from all of the text, all the images that have been put out there already, and then creating a statistical average, if you will, of what you have kind of prompted for. And you can adjust it to a great degree. A lot of the prompts that I have shown you, I've maybe put in a couple different layers of prompts in order to get what I want. But we won't dwell on this, but there are some questions around the ownership of these data, and who is benefiting from it. So, I'm currently paying OpenAI $10 a month for my subscription, and they are obviously deriving that benefit. And the folks that created the data, so the artists who created all those data that it trained on are receiving zilch right now. So I think that these are some questions around ownership that we all need to think through. Just very briefly, when we're thinking about what we're training on, I think it's important to remember that these images that to us are basal cells and melanomas, they are, they're just pixels. And these pixels are basically matrices of numbers that are next to each other. And when it's a color image, it's just three different matrices that are superimposed with different numbers of different intensities. So the idea, you might hear a lot about neural networks. These are not a new idea. These are basically just a math equation that's been around since the 1960s, and where it's modeled on the way that a neuron works. So the inputs are the dendrites, and then they get summed up. And if it gets above a certain threshold function, it'll fire and produce an output. These can be stacked on top of one another to capture more complex relationships in an image or in a data set. And then it will adjust the weights that each one of these inputs is being modeled by, is being adjusted by and modified by in order to maximize the number of answers that it's getting right. So it's like a series of logistic regressions all stacked on top of one another. But I mean, Arnold said it back in the 90s, right? My CPU is a neural net processor, a learning computer. So it goes back a while. So again, this is just an example of how you're stacking all of these neural networks on top of each other in order to capture increasingly complex relationships. And when you hear the term a convolutional neural network, so it's basically taking advantage of linear algebra and multiplying that matrix that you're seeing, that set of pixels by a filter, and then capturing the relationship of those pixels to each other. And the apps that you're using, if you wanna make yourself look extra cute, or if you wanna sharpen an image, those are often using filters as well, kind of of this sort. And you can see some examples here on the right hand side of things that can be used to sharpen or blur an image. But so it'll capture those relationships of those pixels. And so it can average out things like detecting the edges. And when stacked in multiple layers, it can actually begin to capture more and more complex relationships. So let's say we want to train a chatbot to help us on our clinic website, in order to make appointments and get information from patients that we can record. So which category of algorithm might be the most helpful for us? So anyone wanna, so I had Slido set up to do these questions, but it didn't work on this computer. So anyone wanna raise their hands for A? B? C? All right, and D, reinforcement learning. All right, so I see a few more for reinforcement learning. So yes, so there's not a specific perfect example of like the optimal example. There's not like a ground truth for how to ask these questions. So an algorithm using reinforcement learning might be the best way to go about this. We'll be discussing data sets a lot more, but these are critical in creating AI that will work well. And publicly available data sets will let the field move forward more effectively. When you're trying to get an algorithm ready, these data sets are kind of parsed into a training set, a validation set, and then finally you separate a group of data for the test set afterwards. So using that training data set, you create all those initial weights, and then you see how well it's performing on that test set afterwards, and then finally you will validate all of those weights and those outputs at the end. So this is an example of how that training set accuracy might increase over time. The longer you run it and the more you're adjusting those weights in between all of the different neurons. But what you might see is that if you're testing it on your test set, over time it might actually start to do worse. So the more you train it, the worse it might do after a certain point. And so that refers to a phenomenon called overfitting. So it's trying to, if you train these algorithms too long, it'll start to capture random noise and some of these other details that are not relevant. And it'll start to think that they are, and so the performance might actually go down. So the data set will lose generalizability both within the data set and beyond if it becomes too trained. So this curve here does not capture the real relationship between the circles and the Xs. All right, so this is a little different. So if an AI algorithm was trained to diagnose some of these common tumors, and then it encounters a Merkel cell carcinoma and it calls it a BCC. So is this an example of spectrum bias? Can you raise your hands? An out of distribution error? All right. An adversarial example? All right, and an interpretability error. All right. So this is a good example of something that is out of distribution. So this is something that was a data, a diagnosis that significantly deviated from the data set. It never has seen this before, and this is one of the significant potential dangers with AI, right? So if it encounters a diagnosis that it has never seen before, it might call it one of these things that it does have access to. And it might say it with a high degree of confidence. So another example might be if an algorithm encounters a pneumonia in Wuhan that it's never seen before, it might just say it's regular pneumonia, and not recognize that it's something unusual because it's never seen it before. We'll discuss bias to a much greater degree, but essentially, when we're talking about in machine learning, it's gonna be a systematic deviation from the actual correct answer in those model outputs. And it can come from almost at any point in the algorithm. So from those initial data sets to how the study was designed, how you analyze it, and then the kind of outputs that's coming out. So here's just an example of a great study from 2019 showing that the surgical markings in an image actually significantly worsen performance of the algorithm. You guys might have heard this example before, but rulers and markers in images tend to be seen in something that's about to be biopsied. So it tends to be something that's probably malignant, or it's more likely to be malignant. So when they actually cropped out these dots, it improved specificity by a significant portion. I think it was 45%. And then another question that we'll talk a lot about is interpretability. And so this is a paper, an article in the New Yorker a few years ago about dermatology and AI. So these algorithms can give you an answer. It's really hard to know why it said it. I would say this isn't quite right by Siddhartha Murkherjee. It'll give you an output. It'll say, oh, I think it's because of this or that. If you query OpenAI and ChatGPT, it'll give you an explanation, but we don't really know if that's why it said what it said. But there have been some exciting recent papers by my colleagues looking at how to address this interpretability question. So really, we wanna focus today on augmented intelligence. The most important thing is how humans and AI are working together, how they're impacting each other, because that's what is really relevant in the real world. There was a great paper by Shandell's group showing what is the optimal way to look at, for a dermatologist to interpret the outputs from the AI. And in these examples, the faulty AI could actually mislead everyone regardless of how good they were at AI. When we're thinking about AI, this is an important list. There's a cycle that we're going through. You really have to think about what problem are you trying to solve and for who? Who is this going to affect? Are you trying to target something for pigmented lesion experts? That's gonna be very different from an algorithm that you're giving to a nurse practitioner, very different from a triage algorithm. You have to think carefully about the potential biases, out-of-distribution errors, and monitor it over time. With that, I want to thank you all, and I will hand it over to my colleague. Thank you.
Video Summary
The video transcript covers an introduction to AI, discussing core concepts such as machine learning, natural language processing, and deep learning. It explains supervised learning, unsupervised learning, and reinforcement learning in detail with practical examples. The speaker emphasizes the importance of data sets, training algorithms, and avoiding overfitting. They highlight issues like spectrum bias, out-of-distribution errors, bias, and interpretability in AI algorithms. The goal is to focus on augmented intelligence, the collaboration between humans and AI, for effective problem-solving. The speaker encourages careful consideration of biases, errors, and continuous monitoring when implementing AI in various healthcare scenarios. The ultimate aim is to ensure that AI benefits both patients and healthcare providers.
Keywords
AI
machine learning
data sets
healthcare scenarios
collaboration
Legal notice
Copyright © 2024 American Academy of Dermatology. All rights reserved.
Reproduction or republication strictly prohibited without prior written permission.
×
Please select your language
1
English