0917 Responsible Artificial Intellligence By: Dr. Yi Zhang

Thank you so much for this, uh, great opportunity to share some of my current studies and also understanding and experience on this topic.

So for, uh, today's seminar, so generally I will share some, uh, like the topic covering like responsible AI, but with some basic concepts of AI, data science, and then I will provide a little bit technicals about bibliometric graphs and also graph learning techniques. So if we still have time, so I will present one case study on how to develop a graph learning model for predicting knowledge recombination. So that's will be an, I would say, real-world, uh, application of using cutting-edge technologies for technology and innovation management case. Okay.

Uh, yeah, uh, I'm currently an associate professor at the School of Computer Science, also the Australian AI Institute. Uh, yep, University of Technology, Sydney. Uh, yeah, my background or my research is quite similar with Professor Chen. So within a cross-disciplinary area, try to using or developing computational models for handling, uh, real-world problems in science, technology, innovation, and policy studies. Good.

Um, generally, so the first section of the talk will share some general or basic concepts of artificial intelligence and data science. So I just want to provide examples at the moment. So there are four concepts, actually, so we call them data, information, knowledge, and intelligence. I'm not sure you know the differences between or among the four concepts or not. I just provide examples here. So if I provide a screen of numbers, numbers here: 20250917. So what's your general feeling about this string of numbers? So actually, if I don't tell you any extra information or we say additional information, so this is just a string of numbers. So that means this is data. Okay.

So then I let you know that, okay, this refers to the date. So then you got the information. It's, oh, this string of numbers refers to September 17th, 2025. So that's information. Okay. So then, um, the knowledge will be something you know beyond this kind of string of numbers. So you will realize, oh, that's early autumn in Taiwan, but it's early spring in Australia. So that's knowledge. You need to combine some extra information or different kinds of a group of information together and create some knowledge. Okay?

And then the intelligence, it's, okay, then you will realize, okay, because it's early autumn, people may not need t-shirts, the short t-shirts anymore. So t-shirts in Taipei may have some discount, but in in Sydney, for example, we will have our summer soon. So, yeah, I would not say any discount attached with the t-shirts. So that's will be some decision support for business analytics or business intelligence. So that's something we call intelligence. Okay.

So then, uh, from that example, you will find the difference among data, a string of numbers; information, a string of numbers but with some specific meanings; and then knowledge, a group of information representing a specific scenario or specific, um, um, um, I would say concept; and then intelligence, um, moving further with several different kinds of knowledge and achieving the decision support purposes. Okay. So that's a very basic concept in in computer science or in AI data science fields, but they build up the, I would say, the basic elements or fundamental elements of those AI or data science, those kinds of cutting-edge modules. Okay.

So now we've got the data, we've got this kind of information. So actually, the most basic concept is data analytics. It's very easy to understand. We try to analyze the data, but sometimes the purpose of data analysis is to discover knowledge. So sometimes we try to, uh, combine the concepts together. So data analytics, we sometimes call them knowledge discovery in data, or, if you know there is a very famous conference in computer science field, we call them KDD, is Knowledge Discovery in Data. Okay. So I would say that's the somehow that's the step one for data analysis or we say for analyzing data.

So traditional, uh, data analytics is actually quite popular like in the 1990s or or 2000s, yeah, around like 20 years ago. So at that time, people tried to get the data because at that time, people had computers, people had databases, they knew how to store data, and then try to analyze the data. So one of the very famous or representative stories or examples is the beer and diapers story in Walmart, because when you know the data scientists tried to analyze the transaction data in Walmart, they noticed that, oh, sometimes or or in many cases, the customers will purchase the two items together. So that's sort of the one of the most famous stories for association analysis. Try to identify the association relationships between two items. But sometimes you need to delve into the story to understand why people will buy the beer and diapers together. Because those young fathers, when they, you know, finish their work and before they return to to to their home, sometimes, you know, the the the young mother will say, "Okay, yeah, buy some diapers." And then when they the young fathers go to the store, they say, "Oh, I can I can buy some beers together." So that's why they find these kinds of associations. That's a very representative story for data. But you can find the clue there. It's we got the, when we got the data, we got several strings of numbers, we try to find some useful information, maybe some useful knowledge, so we try to discover knowledge from the data. But traditionally, we have two types of studies: one is description, we try to describe data. So, like the beer and diaper story, actually is a is a descriptive study or descriptive statistics, we try to find associations among the two items. But more interesting cases are for prediction. People try to predict, like predict the weathers, predict that the stock market, does how the stock will change in the near future, or, yeah, whatever you can try you want to, you know, foresee in the near future. So that's the prediction study. So that's the clue or that's the point where data analytics can be combined with machine learning. Okay.

Okay, I will introduce machine learning later. You may know that concept, at least. So there are several new issues, but all those kinds of new things are based on the traditional data analytics. One we call them scalability, because we got more and more data. You may have heard the term big data, right? So when we got more and more data, we need to figure out how to efficiently analyze data. Okay. So then we become the big data analytics, and we also want to do some automation that will transfer or that will combine data analytics with some some hot core or hardware, uh, uh, engineering side things. So that will become the automation. But sometimes you will hear similar names like data mining, like data analytics, data analysis. But I would say all of them are sort of the same. Okay. So that's for the step one, data analytics.

But when we try to analyze the data, we will figure out, oh, there are so many types of data. So at that time, we usually classify the types of data into two types. One is we call them structured data, and the other we call them unstructured data. Uh, yeah, there are so many details here. I ignore the, I would not read out read them out, but you could just remember the examples here. If you find the data, for example, in a database or in, for example, Excel, that's structured data. So that is to say, you can put the data in a table, you know what the meaning of each row, you know which, you know the meaning of each column. So that's structured data. They have structures. Okay.

So then for the other data, for example, our videos, our recordings, if you open TikTok, if you open Instagram, this kind of, you know, the the video things or some pictures, all of these kinds of things, you cannot say, okay, that's a row, that's a column, because they don't have these kinds of things. So that's unstructured data. So, nowadays, more and more data are unstructured. So you need to consider how can we transform this kind of unstructured data to either structured data or in any specific form we can relatively easily analyze them. Okay. So that's for the data types.

So then when we find more and more unstructured data, we need to consider, yeah, whether we can find some common patterns to support this kind of data analytics, because if if we can do this kind of things once, whether we can do them 100 times. So the motivation of machine learning is we want to have some methodologies to automate an optimal model building. So that's the, that's the concept here. Generally, it's we can do them once, then we can repeat them 100 times, right? Okay.

So, um, for regarding the concept of machine learning, what I want to emphasize here is the first one, you will find the term "learn," so that is to say, they need to learn something from data. And that something here could be patterns, could be relationships, could be whatever you want to find the clues from the data that could be knowledge, but not that mutual or or, you know, clearly defined knowledge or some in the early stage of the knowledge. So then we call them patterns or or relationships. So, so that's for the two, I would say, the general concepts of machine learning. And the last one is why we need to do machine learning. We want to get this kind of knowledge or patterns or relationships to support decision-making for human beings. We want to achieve automation. Right? So then you can read the second bullet point about the general, uh, definition of machine learning. But like what I described, three key points: one, learn something from data. The second one, the "something" here refers to patterns, relationships, or knowledge. And the last part is the purpose of machine learning is for decision support. Okay.

Um, yep. So that's for general machine learning. Uh, there are some historical stories about machine learning, but the thing you need to remember is machine learning is not, is totally not a new thing. When I think we have another slide on, okay, not here. So when we trace back the history of machine learning, sometimes we consider it's at least have like 50 to 70 years history. So that is to say, just after the Second World War, the computer scientists or or, you know, mathematical researchers or engineers, they tried to develop some algorithms to achieve this kind of learn data, get patterns, and support decision-making. So at that time, neural networks were one of the very popular examples around that time. So neural networks were developed in around the 1970s. So they are not new. Okay.

So even, you know, deep learning, so the the fundamental technical backbone of deep learning is neural networks, actually, but neural networks are totally not new. But due to some reasons, I will explain that later, they were not that active in computer science fields for decades, several decades, and then, yeah, sort of triggered by several, you know, key elements or key technological innovations, and then got the fundamental techniques for the novel deep learning techniques. Okay. So, um, yep, that's the general things for machine learning.

So, sort of the last concept before we start to AI is data science. So you may have heard the terms frequently, right? So data science, you could consider everything or anything related to data is data science. So that is to say, we have a discipline, we have a field, we have a science to to to understand or to analyze the data. So anything related to data, how to understand data, how to get data, how to store data, how to analyze data, how to visualize data, whatever you can consider anything related to data belongs to the field of data science. So that's one thing. The second thing is, machine learning. What's the, what's the relationship between machine learning and data science? So machine learning, if you remember the concept I introduced, is a methodology for analyzing data. So machine learning also belongs to data science, but you will find there are heavy overlaps between data science and AI. I will introduce that later. Okay.

So, now is AI. So if we sort of wrap up for the previous several concepts, we introduced data, we introduced how to analyze data, that's machine learning or traditional data analytics, and then we got anything related to data is data science. So that is to say, from the data part of disciplines, methodologies to handle data, but in the methodology side, in the machine learning side, actually, we have another relative metro discipline that's AI. Okay. So you could see that one AI represents technologies or methodologies, but for analyzing data, like you could see building smart machines capable of performing tasks that require human intelligence. A little bit complicated here, but I try to highlight things. It's for AI, the AI here, it's the, if you remember data analytics, the two, the terms here, the two words, the "analytics" is something related to AI. So we got the data, we tried to get a set of methodologies to analyze data, so the analytics. But the core of AI, the, the, so then you could realize the purpose of AI is to analyze data. And the how say, the the the the final stage of AI is we want to make sure this kind of analytics or this kind of machines can act like human beings' intelligence. So this kind of intelligence is not human intelligence, it's the machine acting like intelligence. That's why we call them artificial, right? Artificial Intelligence. Okay.

Um, there is a little bit of history here. You could find the neural networks. So what I want to add, you know, besides the history of neural networks here, it's actually in in our field, my field and Professor Chan's field, we do technology management. There is a term we call "sleeping beauty." So neural networks are actually a very good example for sleeping beauty. Sleeping beauty, if you know the story, right? There is a, the princess slept for a long time and the kiss and awakened up by the prince. So neural networks were developed in the 1970s, but due to the infrastructural limitations, we don't have that kind of capability to analyze the data, store the data, so we could not release the the maximize the power of neural networks. So at that time, in the 1990s, 1980s, at that time, neural networks were somehow just an algorithm, cannot have good performance, people don't understand how it performs, it cannot solve some good problems, so it's just like a sleeping beauty, sleep, but not die, just a sleep. So then when we come to like 2010s, we got good GPUs, we got, you know, large data storage capacities, and then we got this kind of princess, we got this kind of prince, and then, okay, we combine the neural network algorithms, we got this kind of hardware infrastructure capacities, and then we get this general technological recombination and achieved the great power of neural networks. That's the starting point of deep neural networks, that's the starting point of deep learning in around early 2010s. And if you know ChatGPT, large language models, the backbone of large language models is actually deep neural networks. So you could see developed in the 2010s or even earlier, and sleep or or just stay there quite quite likely for several decades, around like four decades, and then triggered by some newly developed infrastructures or data storage capacities and released the power after 2010s and even after 2020s, right? So, okay. So but why I repeat the story of neural networks, it's why, you know, AI, why everyone talks about AI, because of the huge capability or potential of neural networks here. Okay.

So, okay, that's the general things about AI, but I provide a little bit more explanation. Um, actually, that's not the definition of computer scientists. That's some definition from the, I would say, the social scientists. They try to link AI and their capabilities, different kinds of capabilities, and then try to define four, actually not types, I would say the four stages of AI. The first one is reactive machines. So just like, how say that, like the automation you develop a machine, the machine can repeat different kinds of things and just repeat. They don't have their own algorithm, they just know how you instruct them, they just repeat the same up case by case. So that's the first stage, reactive machine. And in the current stage, we have machine learning algorithms. We can, uh, train our models and ask them to learn some patterns, and then they can, based on the patterns, to recognize some new patterns or or handle some new data. So that's just they have limited memory. So they can recognize the knowledge here. The knowledge refers to or can be linked with this kind of limited memory, and then they can, based on this kind of limited memory, conduct some new activities or new actions beyond something we provided in the original instruction. So that is to say, they can move a little bit forward of our, you know, instructions. Okay. So that's the the the current stage. And then for the, okay, in here we say that's next generation, but I would say that might be our large language models have somehow entered the third stage, that is to say, at that at that stage, AI will have their theory of mind. So sometimes they, they cannot clearly identify themselves. They, they don't know who they are, but they can simulate some scenarios or consequences and then provide a little bit of explainability. So large language models are really a good example here. They can understand your questions. They can, based on the question, create a scenario that, like, okay, you, for example, you try to ask GPT, for example, "Can you provide me a plan for a one-day tour in Taipei?" For example. So they will provide the consequences or the knowledge base of like which kind of, you know, places in Taipei you can visit and and then the distance among those kinds of places, right? And maybe some other things like whether it's like like the prices or or some restaurant combination or some some hiking routines, this kind of these kinds of things. So that's the consequences. Okay. And then based on this kind of consequences, provide several actions, and they can also provide explainability like why I provide you this kind of one-day tour like that way, or they say, if you prefer less hiking, more, you know, more relaxed, relaxed tourism, whatever these kinds of things, I can provide another option. So they can provide this kind of explainability. In computer science, we sometimes call them chain of thought. You may have heard the term before. They provide a how they're thinking or how they're analyzing this kind of data and provide or or digest these kinds of consequences step by step. So we call them chain of thought. So that's the beginning or the starting point of the explainability, but that's the theory of mind. So I would say that's our current, not the next generation, or at least our large language models have already entered that stage. Okay.

And then we go to the, we call them self-awareness. So that's usually the final stage. If you say, I, I, okay, I was the, advisor of some prime ministry, primary, millionaire schools for, you know, those year six students, they need to do some exhibition, they need to do a small case study about AI, understand the impact of AI. So usually for those year six students, their question or common question from them is, "Will we fight with AI?" So I don't know, actually, but I don't think so. But generally, if if there this kind of scenario becomes true, that's just AI will be self-aware. They can know who they are. They know that's the AI. There might be allies, that's human, they're enemies, right? So they need to understand this kind of this kind of situation. They need to be aware of themselves, know have the emotion, have this kind of judgment, and can make decisions by themselves. So that's the future stage. Okay. At the moment, I don't think so. Uh, I don't know, I don't think we will fight with AI, but I think AI will eventually achieve that stage. But before AI eventually achieves or arrives that stage, we are providing different kinds of instructions or or protocols or guidance for computer scientists, but social scientists or end-users to responsibly use AI. That's sort of the starting point of responsible AI. Actually, we will provide more details later, but that's why because when large language models were released and and being engaged in more and more cases, people realized that there are so many potential risks, maybe still opportunities, but risks or challenges we need to handle them or contain them in a controllable way. So, yeah, and the original or the motivation of this kind of responsible AI is the, I would say, the final goal or final stage of AI's development is this kind of self-awareness. Okay. Good.

So, um, actually, I have already introduced what is AI in previous slides. Uh, but I just tried to highlight three verbs here. You could find the the three, I'm not sure you can see my mouse or not, but three verbs here. The first one, perceive. Perceive is AI needs to understand data, understand the environment. So that is to say, they need to have the capability to analyze, to understand the data. Okay. And, yep, we we mentioned that many times, they need to have the capability to learn patterns, learn knowledge. And another thing, maximize. You could understand, you could think about the maximization or maximize things. Like if you ask a large language model a question, they need to provide the best answers, right? How to provide the best answers is a way to maximize. You can see the sentence here, maximize their chance to achieving to achieve the defined goals. Okay. So what's AI? Remember the three words: perceive, learn, and maximize or optimize. Optimize. Okay.

Okay. So another thing I try to compare AI with several traditional concepts, a little bit. Automation and information system. I think relatively easy to understand. If you remember the first stage of AI, they are just reactive machines, that's automation. They can, they may maybe they can learn data or they can perceive data, but they cannot learn any patterns and cannot reproduce that kind of patterns. Okay. So that's automation could be the very early stage of data analytics or very early stage if we put this kind of AI's concept in a relatively broad perspective. So that kind of automation is in a very early stage. Information system is similar. Um, like, you know, when when I was an undergraduate student, at that time, information system actually is a database. You need to use that system to store data, to analyze data, that's all. Okay. So that still still stays on the perceiving data stage. Okay. So that's for information systems.

Uh, yeah, we introduced data analytics, or but we add them as traditional business intelligence or business or no traditional business intelligence, business analytics or data analytics. For that part, we highlight data analysis, but we usually not learn patterns. But the modern data analytics starts to learn patterns. That's why I add "traditional" here. But even the modern data analytics, they just try to identify patterns. But I would say at least they just stay in the second stage. If you remember, I can return the page, they just have the limited memory to learn some very general, general or basic patterns, but they don't have their theory of mind. Okay. To that for the traditional or even not that modern data analytics. Okay.

Then, good example for AI, I have already given that case, ChatGPT. Okay. So I will not spend more time there. So how AI can help us, actually, I just highlight two concepts here. One we call them human-centric AI. Actually, that's not an old concept, and that's a concept I think released by the European Union, 2000, I think 2021 or 2022, around that time, after COVID. So at that time, AI became more and more popular, but the relationship between AI and human beings, we still highlight like human is at the center stage of this kind of scenario, and AI needs to support us. So referring to AI's concepts here, one is we need to use AI to perceive data, to understand something, to summarize something. So one example is sometimes you can, that's not a good example, actually. If you can, you can send upload a paper to ChatGPT, say, "Can you just write down one 100 words for me to describe what the paper talks about?" Right? So then usually, or you used to use one hour or two hours to finish reading the paper, but now you just need to use 10 minutes or five minutes to read that 100-word paragraph. The good thing is efficiency. The bad thing is you can only generally understand the paper but filing all those kind of details. Right? But that's for the summarization. Another thing is, yeah, AI can learn patterns and do this kind of prediction. Very good example is for Instagram. You know, if you click like to one post, or or Facebook, or or TikTok, you click one like here. So the system will recommend more and more similar videos or photos to you because that's based on your preference. So that's the recommender system. This kind of recommendation actually is a prediction. They try to predict your preference based on your previous or historical track record. Okay. So that's is that they try to learn the patterns and make the decision, make the prediction. So we sometimes you also use the term decision support system or, yeah, decision support. So that's a pattern learning behavior or pattern learning activity. So all these kinds of things, we we put them within the general aspect of human-centric AI because the role or the relationship between AI and human beings is AI just to support human beings to conduct something.

So what we want to achieve in the next stage is we call them human and AI collaboration. Actually, Professor Chen introduced, yeah, your next guest seminar, the brain-computer interface part, that's a very good example for human and AI collaboration. I just to provide a little pre-knowledge there. So here, so in that case, let's just say how this kind of how the AI can support human beings to do something. Because for example, if you got the drones, drone can do some deliveries, right? No need to climb the, you know, take the elevators from level one to level 30. You can just use the drones to took all your deliver, your your, you know, takeaway food or whatever. So that is to say, AI supports humans to do something. But at that time, if humans provide sufficient data, like your brain signals, like your previous track records, previous behavioral data, we can use the data to do some real-time analysis, to train, to, you know, do some real real-time training for the AI model to to to to keep enhancing their capability in this kind of prediction, this kind of summarization, or in this kind of, you know, action implementation things. So that is to say, human and AI try to collaborate with each other to enhance each side's capabilities. We enhance AI's learning capabilities, and AI, based on the learning capabilities, provide better services, okay, to for human beings. So that's the human-AI collaboration scenario.

And another example is very common things, or not a common thing, but a common example in computer science field or data science field is we say the current data analytics or the AI, it's we just know or they just know what we have already known. So the example for ChatGPT, it's ChatGPT was trained by the entire, actually, we don't know how OpenAI trained ChatGPT because they don't provide the full details of the training data. But generally, we just use the entire human history's data to train large language models. That is to say, all those kinds of knowledge have already been there, just you don't know that. But that means the entire human beings do not know that. Okay. So that means they, large, the AI or large language models just let you know some knowledge, some knowledge has already been there, but you just not know them. Okay. But what we want to achieve is AI can know something the entire human beings don't know. Don't know. Actually, a very good example for this, I think, is this year's Nobel Prize, last year's Nobel Prize, like we can use AI to find some proteins, right? That's for the Nobel, I think, biology, but awarded to a computer scientist. So that is to say, we try to use AI to discover something we don't know. So that's the challenge. That's something we call inference or reasoning capability. Okay. So that's why I say large language models start the third stage of AI's, you know, the new new age or new a new stage. Yeah. Okay.

But another thing is, um, at the moment, you know, ChatGPT has some issues for the hallucination. We also use the term, we say sometimes they need to, that they will manipulate the results. They will create some false results or fake news, fake misinformation things. Okay. So that is AI still have the, still need to be more effectively maximized, if you remember the term, the verb, effectively maximize their capability in answering the questions. I think I... Okay, so I have a very short video. They just use a very interesting way to generally wrap up what I introduced in the past, like 14 minutes. Okay, just in several seconds. Can you hear the voice?

>> Picture this.

>> Can you hear the voice of the video?

>> That could organize your cup.

>> Yes. Okay, good. Good.

>> Just as you like it. Or serve every member of the house a customized cup of coffee. Makes your day easier, doesn't it? These are the products of artificial intelligence. But why use the term artificial intelligence? Well, these machines are artificially incorporated with humanlike intelligence to perform tasks as we do. This intelligence is built using complex algorithms and mathematical functions. But AI may not be as obvious as in the previous examples. In fact, AI is used in smartphones, cars, social media feeds, video games, banking, surveillance, and many other aspects of our daily life. The real question is, what does an AI do at its core? Here is a robot we built in our lab, which is now dropped onto a field. In spite of a variation in lighting, landscape, and dimensions of the field, the AI robot must perform as expected. This ability to react appropriately to a new situation is called generalized learning. The robot is now at a crossroad. One that is paved and the other rocky. The robot must determine which path to take based on the circumstances. This portrays the robot's reasoning ability. After a short stroll, the robot now encounters a stream that it cannot swim across. Using the plank provided as an input, the robot is able to cross this stream. So our robot uses the given input and finds the solution for a problem. This is problem solving. These three capabilities make the robot artificially intelligent. In short, AI provides machines with the capability to adapt, reason, and provide solutions. Well, now that we know what AI is, let's have a look at the two broad categories an AI is classified into. Weak AI, also called narrow AI, focuses solely on one task. For example, AlphaGo is a maestro of the game Go, but you can't expect it to be even remotely good at chess. This makes AlphaGo a weak AI. You might say Alexa is definitely not a weak AI since it can perform multiple tasks. Well, that's not really true. When you ask Alexa to play Despacito, it picks up the key words "play" and "Despacito" and runs a program it is trained to. Alexa cannot respond to a question it isn't trained to answer. For instance, try asking Alexa the status of traffic from work to home. Alexa cannot provide you this information as she is not trained to. And that brings us to our second category of AI, strong AI. Now, this is much like the robots that only exist in fiction as of now. Ultron from Avengers is an ideal example of a strong AI. That's because it's self-aware and eventually even develops emotions. This makes the AI's response unpredictable. You must be wondering, well, how is artificial intelligence different from machine learning and deep learning? We saw what AI is. Machine learning is a technique to achieve AI, and deep learning, in turn, is a subset of machine learning. Machine learning provides a machine with the capability to learn from data and experience through algorithms. Deep learning does this learning through ways inspired by the human brain. This means through deep learning, data and patterns can be better perceived. Ray Kurzweil, a well-known futurist, predicts that by the year 2045, we would have robots as smart as humans. This is called the point of singularity. Well, that's not all. In fact, Elon Musk predicts that the human mind and body will be enhanced by AI implants, which would make us partly cyborgs. So here's a question for you. Which of the below AI projects don't exist yet? A an AI robot with citizen.

Okay, I will stop here considering the time, but generally, they cover the main points of what I introduced in the first section, which is about the concepts and a little bit of interaction with the AI technical things and also the real-world examples or cases. Okay, good. Let me move to the next slides.

Okay, sort of the wrap-up, I will open word these kinds of things. Actually, I mentioned a little bit, what is the relationship between AI and humans? So actually nowadays, I still consider people will stay in the like tools and operators. No, before that, it's tools and operators, but nowadays it's like assistants and directors. What's the difference between that one? It's if we consider you just use Google, okay, like 10 years ago or five years ago, before large models, you just ask questions or search or input some like keywords there, they just provide the answers, right? They are tools and you are the operators, just try to find the information there. They can, they may have some learning capabilities, but you still need to spend a lot of time to figure out, okay, that's the best search result returned by Google, but that second one maybe not the third one is good, then you need to do some combination. Okay, so that's the stage of AI before large models. But nowadays, why I say is assistants and directors, because AI has the capabilities to generalization, to summarization things. So that means they can search 10 results and based on their criteria, based on their judgments, they try to summarize them into one paragraph, just like your assistant, right? They, you, you ask them to read 10 papers and then they told you, okay, in 100 words. But the eventually, we would say the relationship between human and AI is collaborators. They support us, and we support them to be more, smarter, right? That's the relationship. Sort of the wrap-up.

Um, yeah, we mentioned quite a lot of concepts actually. In the video, they also provide an introduction, the, you know, the differentiation between deep learning, machine learning, and artificial intelligence. But we also mentioned data analytics, data science. So you could see anything related to data is in data science. So that's why the data part is also within the field of AI, because for AI, if you want to train a model, train algorithms, you need to have data. So machine learning, why also within the part of data science? Because you need to use the algorithms to analyze data, right? So that is to say, the algorithms for analyzing data also belong to data science. That's why you can see data science and AI are overlapped with each other. Okay. So that's for the concept level.

So the last one, actually, I mentioned a little bit for the large language models, but here I just tried to provide a little bit more details like what is a large language model and why we call them large language models. Okay. Okay. Okay. The first one, language model. Why we call them language model in general? Ignore those kind of huge paragraphs, not huge, the paragraphs here. So for the language model, because the key purpose or the motivation of those kinds of methodologies is to make models to understand or to analyze human language. Okay. So that's the main purpose. The initial motivation is to analyze. The second one, uh, uh, to understand. And this eventually, the eventual or the last stage, the final stage, is they can generate human language. Okay. So that's why we call them language model. So anything or the core data about this kind of models actually are human language. But this kind of human language could be written down in in some records, in texts. They can also be recorded in videos or or in some audio, the the the speech, or they can also be in some image, because sometimes the pictures also a type of human language. Okay. So that's the first thing why we call them language models.

And the general idea of the kind of the methodology or the core methodology behind language models actually is, you could see that one, the second one actually is a probabilistic model. We just try to analyze the distribution of those kinds of words in human language. And something just like, if you say, how the possibility people will use the words after "how," for example, people, the language model will say, okay, 80% people will use "are" after "how," because we will say "how are you?" And another 20% people will say "how is something?" And, okay, another 1% is "how I don't know." Another maybe a bad example, whatever. So they just use this kind of statistical way to find the connections or probability between the pairwise words or pairwise sentences or pairwise terms and use that way to predict what the next word. So that's a core concept of language model. You may know that one, but that will help you understand why language models can generate answers to your questions, because they just use a mathematical way or statistical way to calculate the probability word by word. Right? How the 80% "are," and the 80% "you," and the 80% is a question mark, and then like, yeah, the following sentences, something like that way. Okay, I will not go deep, but just like this way, just remember it's a probabilistic model. Okay. So that's for the, uh, language model.

So but they don't count models, they don't count words or terms in the traditional way like how we write it down in your notebook. For example, if you want to write that five terms: women, child, men, queen, king. We just write it down the five terms, right? But in computer science, we somehow will store them in a vector. Okay? So you can imagine each feature or each cell here may have their specific meaning. Like if we have say, that's for the royalty, that's for like how the muscle, the the scale of the muscle, how they look like a female, the age. So different kinds of meaning of the the feature. So we'll say, okay, for king, the royalty is really high, the muscle could be high, and the female side is quite low, and king might be not that young, a little bit old, right? So we can have a similar way to quantify for queen, women, and princess. Okay. So once we got this kind of vectors, so then all this kind of relationship among those terms will become mathematics. So that's something we call word vectors. So the example is, you could imagine if we have a vector of king, then we minus the vector of man, what will be left? We have a king vector. We have a man. Oh, we don't have, I have princess vectors. Actually, the the remaining information could be the royalty. Okay. And then, oops. Yeah, the royalty could be the royalty. Okay. And then we add, then the women vector. Okay. So that's we have the royalty remaining with king minus man. And then we add the feature of the women's vector. So that will be the vector for queen. So that is to say, in the mathematical way, we can use king minus men and plus women, then we will get a vector for queen. So that's how in computer science field we do this kind of calculation and use the mathematical or statistical or probabilistic way to do this kind of calculation. Okay.

And then, so we move things further. We use this kind of mathematical way to represent words. We try to find the probability word by word. So that is to say, then you could imagine if we got the probability, that is to say, we got the capability for prediction. So this capability is generation. That's why sometimes we call them generative AI or generalized language model. So that's why we call them generalization, because once you got one word, and based on your huge knowledge on the entire human knowledge or human history's data, you can clearly know what's the highest probability word after "how" this term, right? Then you can write the entire paragraph, entire paper. So that's how the language model can generate content. Okay.

Okay. That's for the model part. Then you need to realize if we can, if if we want a model to be equipped with this kind of capability, how large the parameters or the features you could consider those kind of features. All those kind of features could be parameters. So then you need to consider how large these kind of features, features or these kind of parameters we need to have. You may have heard or you may not. So we have a BERT model, actually, they are, they are language models, but comparably, we call them small language models, but they still have 29 million parameters. So if you have some basic knowledge about like K-means clustering algorithm or or how do you say that, like PCA, the principal component analysis, I would say the parameters you need to figure out around like less than five, right? So but here we have 29 million parameters. And the large version of BERT, they have more than 300 million parameters. But compared to large language models, we still call them small language models. Okay. So why we call them large ambient models? Because the parameter level are large. They have too many or so many parameters. I think we have a, yeah. Okay. I just provide the information. Oh, no. Oh, I don't have the number. Okay. Okay. Anyway, so for large language models, for like like like GPT or if you know Llama, they usually, the smallest version of the large language model, the smallest version of GPT, they have around like several billion parameters. That's the smallest version. And the large version, they have several hundred billion parameters. So that is to say, if you want your large language model smarter, you need to have more parameters. That's why we need, you know, more and more powerful computers or more and more powerful servers. Sometimes people just say, yeah, you cannot run large or fine-tune, you cannot fine-tune large models on your laptop, you need to have a server. So that's because they have too many parameters to handle, to store, to be analyzed. Okay. So that's the general thing about, um, language model and large language model. Okay.

And then, yep, the very good example why we call them GPT is Generative Pre-trained Transformer. Actually, the three terms cover the core technical concepts of large language models. Generative, if you remember that one, they have, they can generate words or language or natural languages word by word, right? The generating languages, that's why we call them generative. And then Pre-trained. We have so many parameters, just to say, you could imagine we have 100 empty cells, you need to figure out, you need to fill in each cell with a number. So how to fill numbers in those kind of cells, that's the training process. But you can never train this kind of billion parameters every time, right? You say, yeah, today I have one question, so I need to train the billion parameters. Next time I have another question, I need to train that again. So we need to pre-train these kind of parameters and make them ready for some general uses, like GPT, so or ChatGPT. So that is to say, they are pre-trained for future use, for future uses. And the Transformer actually is a very modern type of deep learning technique. So it's a new architecture, but the core is deep neural networks. So that's why we call them GPT, short for Generative Pre-trained Transformer. And ChatGPT, we add "chat" here because they can have very good performance in generating language, that is to say, they can chat with you. So that's why they put "chat" before GPT, the ChatGPT. Okay. There are also some other representative large language models, but the major products are from

The United States, CH GPT, Llama, or Cloud, or Grok, or only the deepse from mainland China. Okay.

Anyway, so that's all the large language motor pot. You sort of wrap up why large language motors are so smart. I just put all those kind of content. So they are pretend models using large data. You could sometimes I just use that example. You could consider a large language model could be a very upgrade or anyway, it's instant. So large models are instant. So they know everything. Okay. So that's why they are so smart.

But why we sometimes in computer science we say we need to fine-tune large language models? The good example here is, for example, you can ask Einstein, "What's the best route for me to drive from Taipei to Hsinchu?" Einstein may not know that because he, I'm not sure whether Einstein has ever been to Taiwan, but anyway, he never drove on that route between the two cities. So he knows everything, but for this specific point, he does not have the knowledge to understand this scenario. So why sometimes in computer science field or in specific application areas, people need to fine-tune large language models? The purpose for fine-tuning large language models is to provide domain-specific knowledge. For example, we provide the entire maps of Taiwan, or at least all routes between Taipei and Kaohsiung or Hsinchu to Einstein, then Einstein can know, "Okay, that's the best path," like Google Maps, right? So that's the first one, pre-trained models.

But sometimes we need to retrain them using domain-specific data. And the second one, why they are large or why they are smart? Because we use billions, sometimes trillions of parameters to control the model. So you could consider it a very precise machine, and that machine has like a billion or trillion buttons you need to click. So that's why they are so smart because they are controlled or they are developed in such a precise way. Okay. So that's the second one.

And the last one, if you remember the four stages of AI, because they are in the theory of mind stage, they try to find patterns, they try to develop their own actions beyond these kind of points. Okay. So that's for large language models.

Um, yeah, some limitations. Um, two things. The first one, if you have ever used large language models, at least ChatGPT, even if you just use their interface to ask questions, you need to pay them, right? Particularly for the advanced models. But for computer scientists or for some large-scale use of large language models, you need to use their APIs, run the APIs through coding or programs. So you need to pay based on tokens, like how many words you have. Then you need to pay OpenAI for using, uh, uh, GPT. So that's, they are costly.

The second one, it's, um, yeah, I provide a little bit more here. Uh, sometimes people say, or every time people argue, complain that large language models will have the hallucination issue because they will provide incorrect messages. The reason behind this kind of hallucination, actually, is they, when computer scientists were training large language models, they are using a wrong or using a using, I would say, a wrong, using a wrong awarding function. Just like if you have a child, you just ask them, "You can never say 'I don't know.'" If the child says "I don't know," you will punish them. And if they say anything, you will award them, say, "Oh, that's good, you're a good boy." So during this kind of training process, large language models learn that "I cannot say 'I don't know,' I cannot say nothing, I need to provide some answers, no matter that answers are right or wrong. Even that's a wrong answer, I can still get the candy." So during this kind of training process, large language models will do this kind of, will have this kind of hallucination issues. Actually, computer scientists are working on that part. Uh, this area we call them hallucination detection. It's a very cutting-edge area in the machine learning field. Okay.

Okay. Um, then that's the main point. After our break is ethical concerns. When are using large language models, when the other AI models, we will have very deep concern about some ethical issues. Which kind of data are used for training? Whether they will provide some, you know, biased recommendations, or this kind of recommendations will be, um, still biased, but biased to, uh, the different kind of demographic attributes. Okay.

Um, yeah, they are, at the moment, they have no self-awareness. Okay. We will stop here and have a 10 minutes break, and then have the rest of the part for the responsibility. Okay.

Okay, let us restart. Actually, I have one video here, but I don't want to run it. It will be a little bit long, like 15 minutes. But when you get the slides, yeah, just feel free to to watch it. I think they are using a very interesting way to provide. You could see the Einstein sitting here. They use a very interesting way to introduce what, what the actually, a little bit technical part also the application part of generative AI. Okay. So I will skip it off. Okay.

Okay. Then we will move to, uh, the AI CS, also the responsible AI things. Um, the general idea of responsible AI or AI CS is when we got AI or when we got large language models as kind of products or anything, or we could consider still considered as a child. Actually, in human, in in the human society, when we have a child, we have a very, uh, precise or very, I would say, a very natural system to teach or educate the child what is right, what is wrong, right? And then this kind of education system can guarantee the, the well-development or of the child, or help them grow up, uh, uh, you know, mentally and physically healthy, something like that. But for AI as a product, at the moment, we don't have this kind of education system, or we don't have the law or regulation to teach the developers, that the computer scientists, the end users, our broad societies, what is right and what is wrong. And in many parts of the tech, particularly the technical details of AI are not transparent to the end users. That's to say, you got the tool, they provide you some help, but you don't know how they produce or how they get this kind of help. Or in a simple word, in a simple word, so you just like you got the money, but you don't know where the money comes from, right? So for those kind of situations or consequences, we need to discuss or or we argue the urgent need of AI governance, or AI regulation, or responsible AI. Okay. So that's sort of the starting point of the our following talks. Okay.

At the beginning of the part, I tried to list several current ethical concerns of AI. So I believe you may have heard some of them here and there, but I tried to provide a little bit more details, maybe not the technical side, but I believe more what it is and and what they are, why we have this kind of concern. Okay.

The first one actually related to fairness, transparency, and explainability. I try to explain them starting from transparency. Um, if you still remember, we mentioned deep learning, deep neural network. The core idea of neural network actually given in that in that YouTube video, they just developed the algorithm to try to make the algorithm act like the human brain. So they try to put those kind of, you could consider that's a network or or whatever, they have different kind of layers, okay, and how those kind of communication between those layers, somehow it's based on the probabilistic models. So once we have the probability, anything will become uncertain, right? So that's, we'll come to the transparency issue. That's to say, you will know 1 + 1 = 2. That's very clear and straightforward. But if I tell you that if I put one and one plus one, then I put in a in a container and I add something, you know, we use another example. I put two kinds of food in a container and I add some spices, I add some chilies, add some sugar, salt, whatever. And then just mix them together, and then I get my final food for you. You will consider the food is delicious, but you will never know how this food was created. You just know those kind of recipes, right? So that's the untransparency issue. Maybe not an exactly good example, but somehow similar with the mechanism of deep neural network. You have the input, you have the output, but you don't know what happened between this process. So that's the transparency. You cannot know this kind of stepwise actions. That's transparency.

And the explainability is, okay, you got your final food. I can try my best to let you know, okay, this food contains rice, or contains, uh, some, some protein, for some, some fat, whatever, these kind of things. So that's the explainability. I try to explain what it is, particularly for the final outcomes. Okay. So, but sometimes even for the explainability, the model cannot achieve them. Okay. For example, in, if you have ever used some traditional or old version of the, like the, the Instagram, for example, um, I think there is a one page you can see different kinds of posts, and I think nowadays somehow, I think it's some time, I think one year or several years before, they provide a short note like, "Explain why they provide that photo or that video to you." Sometimes they will say, "Oh, someone you followed liked the video, or you liked similar videos before." That's the explainability. Okay. So we need to explain why we provide this suggestion, why we provide this kind of prediction. So that's the explainability issue.

Then it comes to the fairness issue. So the fairness means whether we consider this kind of results will not create any discrimination or bias to anyone. One of the most famous examples here is many years ago, I think, um, I think it's Amazon. Amazon tried to train a model that can do the initial check of résumés because they received quite a lot of, you know, the résumés from the job applicants. So they tried to, not they tried, they collected the data from the successful middle-level managers' information in the 1990s, and they used that data to train their résumé filing model, whatever this kind of model. And then they found a very strong discrimination to female and, you know, the female applicants. That's one. Another one is any applicants not, how do you say that, um, any applicants not from, not from the white background, like black people, like like the Asian ethnicity, even for the, you know, the, the, like from the Spanish, this kind of background, you will not be, you cannot pass this kind of initial check. The reason is the training data from the 1990s, Amazon's, uh, the system, the majority of the successful middle-level managers are male and white. So that will become some very core features in that machine learning model, the, the résumé filter model. Then they will create very serious discrimination to those kind of specific demographic features. So that's the fairness. Okay.

So then it comes to the conclusion, like you can have a very smart model, and the model can provide you some decisions, can provide you some suggestions. Whether these suggestions are fair, that's one thing. And another thing is sometimes, if you got the results, or I provide the, the example here, have you ever searched your name? Maybe you, if you, okay, like your supervisor's name, or some, not that famous people's name in GPT. Sometimes, like the hallucination issue, they will create a, you know, not bad and similar portfolio for that person, but that's not this person. So I tried to search my name in ChatGPT several years ago, just when, uh, the, the ChatGPT was launched. So they say I'm a female researcher. I am the dean of my faculty. No, but they say my research background is in something between information system and AI. Maybe, right? Yeah. They just mix of correct information and fake information and generate the results. They cannot explain where they got the message, and whether the message is wrong, and whether they create or manipulate or, you know, or just, yeah, generate this kind of hallucination data. So that's the first part. I would say that's one of the most serious issues. They usually relate relate with each other because the model is not transparent. Somehow we don't, we cannot guarantee they will generate fair results because they cannot provide sufficient explainability for the final results. Sometimes they will be biased to some specific features. Okay. So that's the first part, fairness, transparency, and explainability.

Actually, I provided one reference here. This is a very interesting paper, uh, led by a very famous researcher, Cassie Suimoto. She is a chair professor in Georgia Tech. Um, yeah, I think Professor Chun knows her. So this study tried to ask GPT to generate some narratives for education purpose for the preliminary schools. So they find a very serious gender bias in those names, in those narratives. For example, some white boys' names always help people with a name in Asia, with a name in in black, or they always try to have some specific name, for example, Mike. This kind of Mike always helping people, never being helped by anyone else. But Mike, obviously, is a male name, right? So they try to do this kind of empirical tests and report the results in that paper. So if you are interested in that part, yeah, check the results in the paper. Okay.

That's the first one. The second one, I think, is easy to understand. They have, uh, whether they can protect your privacy, whether they can introduce some security issues. One thing is, I think I'm not sure you have the similar feeling or not. You use, we use various kinds of social media platforms or apps in our phone. Sometime, for example, if you chat with your friends or call with your friends in, in, I know in Taiwan, you use Line, right? You use Line, and then you go to TikTok, for example, then you find, "Whoa, TikTok provides you some recommendations just related to the content you just talked with your friends in Line." The reason behind that one is, in, in, in laptops or in mobile phones, they have a common place to store something we call them cookies. So that's some temporary data. So they will be automatically, uh, delete, uh, regularly or or within some time frame. So everyone, if you have the authority, you can use that part of the data. Even nowadays, you know, different kinds of apps, they provide different kinds of terms and agreements, say, or, you know, informing you how they will use or how they can use the data in your mobile phone, in your laptop, but you know, you will never know how they use the data. So maybe they just use them, I store them, or transfer them, but they just don't let you know. So that's the security issue or privacy issue.

And another thing is, computer vision is a very popular field in AI, but they will have very serious ethical issues, privacy issues about the surveillance system, right? This kind of surveillance system could be very helpful for the national security because, you know, they can record you appear in that building and disappear at the end of that corner. Something they can, based on these kind of behaviors, to track where you are, who you are, and some potential, you know, risks, these kind of things. But, yeah, whether these kind of things are really ethical or responsible, whether they can protect your privacy, somehow you don't yet, you just don't want anyone to know where you are, right? So, yeah, this kind of privacy issue. Actually, another thing, I'm not sure you have ever realized or not. Any questions you asked, you ask ChatGPT, any documents you uploaded to ChatGPT can be, or maybe, or maybe have already been used for training ChatGPT. So that's to say, you somehow you can choose the option, say, "You don't want to use the data to train your photo," but if you don't delete, if you don't delete that record, and ask similar questions, you ask, keep asking some questions, you will find GPT will just get some answers from the information, uh, in your questions asked in some other, you know, transactions or in other records. So that's something we call data leakage, but also related to your privacy protection issue. So that's why, you know, in my university in Australia, we don't have a very clear or, or, you know, public announcement say, "You cannot use ChatGPT for some academic research use," but we have some common sense, like if you are trying to write a proposal, for example, research proposal, which is a novel idea, never been published before, or some confidential information related to, for example, your national security or military projects, you cannot use large language models because you never know how they will use that data. So that's for the privacy and and security issue. Actually, the example provided here also related to the example I, uh, just shared. Okay.

Okay. Good. So another one, the, uh, contestability. Actually, it's also related to the, the, I would say, the accountability. I try to talk them together. So the contestability, actually, is a common question. I was asked it by like someone like my parents, like my taxi, like the taxi drivers, you know, these kind of people, because everyone are a little bit curious or a a little bit cautious, sometimes just a fear whether AI will take the place of human beings, particularly in, in some specific jobs, right? People will say, "Yeah, if we have autonomous cars, we don't need drivers anymore. If we somehow, if AI can code, we don't need computer scientists anymore." So that's the contestability between AI and humans, right? So whether they are really have this kind of a competition situation. If you recall what we talked about human-AI collaboration, whether the based on our rules, our regulations, our guidelines, they can become friends with each other rather than competitors, okay? So that's for the, uh, contestability.

And then the accountability. It's in such kind of scenario, human and AI collaboration, who will take the responsibility for potential risks or for some errors, mistakes, right? The example here, it's that might be a bad example. If a self-driving car hit someone, who will take the responsibility of this kind of accident? Either the developer, for the algorithm, for the AI, or the owner of the car, or the passenger of the car, or the, the, the company of the car. Yeah. Who will take the responsibility? Actually, I don't see any laws or or regulations about this kind of case. Maybe not become true, but I know in, in mainland China, in many cities, they have the examination zones for self-driving cars. And when I was in, in Atlanta, that's the last time where I met a professor, Chinese, in person. I just saw some self-driving cars in Atlanta. So that's to say, United States, both United States and China have these kind of examinations, just in, you know, in our daily life, you may know, you may not know them, okay? So who will take the responsibility when this kind of accidents happened, right? Okay.

So then is reliability and the safety. So, yeah, I provide a question here. Do you trust ChatGPT? So nowadays, like what I say, if you still remember, we say it's assistant directors, right? You, you make the decision based on the answers, questions provided by ChatGPT. But if in that kind of real-world scenario, AI or robots were handling some specific jobs, can we really trust them? Right? And another example, like the deepfake. If you know that one, the fake photos or videos created by AI. How can you recognize that's the deepfake videos, deepfake photos, and that's the real things, right? The ChatGPT can, can, you know, simulate different types of wars. And if you use different kind of training models, uh, to train GPT, and they can, you know, oops, sorry for the, okay. You can still, uh, okay. Let that example. I just watched some videos, just like Obama or Trump are talking, but actually these kind of videos are created by AI, right? So how can we, just like our general public, how can they recognize that's AI, that's not AI? So that could be another issue for the reliability and safety, right? Because I provide a video, say, "Oh, that's some video from Obama, that's some video from your parents, that's some video from your children." That will be some very serious issue if that kind of things were created by AI, right? Okay. So that's for the reliability and safety. Okay.

Um, then it comes to some a little bit confusing concepts. Um, in computer science field or the broad social science field, there are generally four core concepts: Explainable AI, relatively easy to understand, Trustworthy AI, Ethical AI, and Responsible AI. So that's the general four core concepts. I provide a little bit, um, explanation here. Um, the first one is quite easy. Explainable AI highlights the explainability things. They, they need to equip AI with the capability to explain the results they generated. Okay. So that's for Explainable AI.

And then the Trustworthy AI is many in the computer science field. They need to consider this kind of AI model is reliable. This kind of reliable could be, we trust their results, and during their data processing or during their data analytics process, they can maintain the security of the data, they can, you know, they can keep the, you know, the, the all those kind of sensitive, keep the sensitive data or private data in a container with several, you know, novel algorithms, and still can provide the final results. Okay. So in that case, um, Trustworthy AI is still a concept within the computer science field. People try to use trustworthy to consider this kind of model is reliable, and then use several measures or indicators to describe or to measure this kind of trustworthiness. Okay. So that's the second concept. You just need to remember, related to reliability and within the field of computer science. Okay.

Uh, Ethical AI, sometimes we also call them AI ethics, is sort of the first concept that comes beyond the AI community or the computer science community. It extends or interacts with broad, like I just mentioned, academia, industry, government, and public care, but also across different kinds of fields. For example, social scientists will understand what is AI's ethical issues. Lawyers will understand whether we can provide AI governance. Actually, AI governance comes after AI ethics because we have this kind of ethical concern, people need to consider how can we reduce or how can we mitigate this kind of ethics. Okay. So AI ethics generally related to the ethical concerns from the AI models, and it comes beyond the computer science community to the broad society. Okay. So that's for the AI ethics.

And for Responsible AI is, yeah, it is beyond the computer science community, but we highlight it's a socio-technical system. We need to incorporate with both the technical side of the model, the models' algorithms, and the practical side of the society and users. So we call them Responsible AI is to highlight this is a socio-technical system with AI techniques. We need to consider the developers, the models, these kind of scientists, and the practitioners, like our end users, general public, the lawyers, the, the policymakers. So that's a somehow this is a responsible AI, and that's to say we need to provide some guidelines for the AI developers to create some responsible AI models, and we also need to guide our AI end users to practice or use AI models in a responsible way. So comparably, the first one is Responsible AI could cover all the three concepts we introduced before this one. And the second one is Responsible AI provides a comprehensive understanding from the AI modeling side, the technical side, to the social impact side, to the social side. Okay. So that's a very high-level understanding of Responsible AI. Okay.

Uh, and then I try to link those kind of ethical concerns, uh, with specific data analytics or we say the AI technologies. I will not read or introduce all of them, but generally, you could see that that's the statistics. Actually, it's from bibliometrics. I, I'm not sure you know the basic concepts or not, but generally, we try to got this kind of information from AI related AI ethics related papers, and we see how people develop or discuss, uh, the two concepts together. So then we find if people are talking about machine learning, machine learning are related to almost all those kind of issues. Could be fairness, if you remember the Amazon's case, it could be the accountability, like the autonomous cars, because the, the, the technical core of autonomous cars is machine learning, and the privacy, whether we can, uh, keep the, the, the, the sensitive data, and we can create still create the, the precise, uh, results. So that's something in, in the computer science field, we call them federated learning. So that's a machine learning side.

And another thing, I provide example here for the, uh, neuroscience. Yeah, actually, yeah, I would be curious about if you ask this kind of question to your next guest speaker, how they can answer or address these kind of concerns, because when you are trying to analyze or access the human brain signals or human brain data, there will be a serious, serious privacy issue. Of course, uh, before you know, they conduct this kind of experiments, all those kind of, the, the end users will sign some agreements or consent to to say, "Yeah, I agree to use my brain signals or what, whatever these kind of things." But when this kind of techniques or these kind of products become the real-world products and be used in your daily life, like your TikTok, like your Instagram, like your Facebook, whatever this kind of Line, these kind of things, it that will become a huge black box because our general end user, you know nothing about the technical side, and the governments or the, the governments provide guidelines, but how to really, uh, regulate or monitor the use of the data in your own cell phones, in your own laptops, sometimes it's really hard. So how can we really practice this kind of privacy protection? How we can really, you know, regulate the use of the brain signals or this kind of private data in those kind of AI techniques or AI products, that could be a challenge. Okay.

I will not, I will skip off the rest examples, but feel free to to find, uh, this kind of things or or pick up several interesting points. Um, I will also not explore more about that one, but generally, it's relatively easy to understand. We try to see which part of the universities they highlight this kind of study. At the moment, most like English-speaking countries have strong interest on this kind of AI CS study. But this one, uh, this study was conducted around 2020, but I believe in the past five years, more and more Asian countries, like Taiwanese researchers or the researchers from mainland China, whatever, Japanese, Korean researchers, I do see a lot of papers published from those kind of regions. Okay. So, yeah, but see that that's sort of some results published, uh, at least five years ago. So that's leading universities from those kind of English-speaking countries, they hold strong interests on those kind of AI ethics studies. Okay.

Okay. Then we delve into the Responsible AI. So generally, we introduced what is AI's ethical concerns? What the general concepts and or related to Responsible AI. So the core idea is, what's, how can we really define this kind of responsibility? And who, like, because we describe them as socio-technical systems, which kind of stakeholders or which kind of, you know, end users or or parties are involved in this kind of Responsible AI's socio-technical system? Uh, the first one is actually the current metro or where recognized responsible AI protocols are produced by industry rather than university or governments. I think it's easy to understand that one. I think one of the most popular one is from Microsoft, and also some of them are from like IBM or some management consultation companies, like KPMG, like McKinsey, these kind of things, because they standing, they are standing at the level of end-users level to oversee the technical development and the users' behaviors or users' concerns of these kind of AI things. They try to identify several protocols related to the AI's development and uses, and then you will find many concepts actually we have already introduced in our AI, our AI's concerns. But generally, you could see that one, the principles are proposed based on those kind of standards that what we have already introduced. But if you delve into those kind of specific reports on the, like the protocols of Responsible AI, they will provide some guidelines for the practice, for the AI developers' practices, like when you are developing the algorithms or when you are modeling, how can you achieve this kind of accountability?

So I add a note here. It's what the real practice from the AI communities, because we spend around around like one hour or yes, around one hour to introduce what is AI, what the deep neural networks, what the large language models, right? So then from the AI community's point of view, which kind of concept or which kind of information related to this kind of AI's responsibility, and from the society point of view, from the end users point of view, how can we practice these kind of principles. So in the following part, I will generally focus on the core technical techniques of the Responsible AI, that's to say, how the AI community addresses this kind of Responsible AI's principles. But for how the, uh, governments, how the general society provide these kind of protocols, I will not cover that part, but I would say that's still in the emerging field. Different kind of countries, even the federal governments and domestic governments have different kind of visions or or protocols. So that's just to say this is still, you know, uh, under development, and different kind of, particularly different kind of regions, actually, like in the European Union, they sort of the pioneer in that field, provide some AI ethics guidance or these kind of things. And like in Australia, people try to, or the government try to follow these kind of protocols, but they were, they still need to consider the context, the context of their own country, right? The Australian context. So they need to modify some part of the contents related to their their own country situation. So slightly different, but the general basic parts are quite similar. But generally, we will, we will not touch that the social part of the Responsible AI, but not that technical. Okay. So don't be afraid of the technical part. So, yeah, generally we do a bibliometric study. Uh, we just looking to around like 18,000 responsible AI related research articles published before 2025, but we focus on the AI related journals and conferences, and, uh, uh, interestingly, yep, um, the largest part, I think it's from China, but USA quite similar, one is around 3,800 records. Another one is 3,700 records. And the other countries, they also produced similar studies or, you know, some appealing studies, but not that comparable to the United States and China. But more interestingly, we generated this kind of treemap for the tech, actually the technical tree of Responsible AI. I'm sorry that the font here is quite small, but yeah, because it's too technical, I just highlight that here. When we try to delve into that, like 18,000 articles, we try to understand what it is, what they are focusing on, and what the core contribution from the AI's point of the AI community's point of view. So then we try to summarize them into the four sort of the sub-areas of AI. One is the from the machine learning point of view, they highlight different kind of algorithms. So like the general idea is how to make the algorithm trustworthy, if you remember the Trustworthy AI. So if the, because the general assumption there is, if the model is reliable enough, we will produce responsible results, and this kind of responsible really depends on the core focus of the modeling. For example, someone just to focus on privacy preserving, someone just to focusing on the, uh, accuracy. That's the trust. That's the trust. Another part of the trust, and also someone just to focus on the fairness, right? So that's the first part, modeling side, uh, from the machine learning community.

And from the data community, they focus on data. So one of the major concern contribution from the data community actually is the privacy. So how can we, you know, protect the privacy when we are collecting data, when we are storing the data, when we are analyzing the data, whether we can, if you remember I mentioned them in machine learning field, we call them federated learning, but in data mining field, they have something plus federated learning. That's to say, like a, like a, you know, the one-stop, one-stop analysis from the beginning to the, the data analysis will never touch the core information of the data, and then still can get the final results. Um, one example, maybe not that accurate, it's if you're looking to, you know, the, like like the browsers, like the Chrome, right? They will store your password, right? So the easiest way to store the password, just say, "Okay, that's your username, that's your password," for example, it's ABC and then 123 and then some, you know, specific symbols. But in a, in a relatively, if we call them privacy-preserving way, is we try to use another way to transform this kind of password into some other format. For example, we can transform A to, uh, the capitalized B, and with one A with the number one, and with one, for example, at the symbol that means A, right? And then we can use some other one, other way to transform B. Actually, in, in computer science, this related to cybersecurity, but this is also related to something in mathematics. So, how do you say that? We actually we call them cryptography, or in Chinese, we call them mimash, okay? Like how to, how to produce the privacy-protected password, how to store this kind of password, okay? Anyway, that's for the, the data side.

And then mathematics. I mentioned a little bit, like how to use your mathematical formulas to transform those kind of passwords. And also for the algorithm side, you also need to contribute from the mathematics or statistics. Okay.

And then from the network side, that's easy to understand. If you need to transform the data, for example, if you are using the cloud data, because cloud data means you will allow your data be stored online or be stored in some other servers rather than your local laptops. So how can we protect the data, particularly when we transforming the data, uploading the data from your local laptops to the clouds? In this data processing or data transforming process, how can we protect the privacy? So, uh, that's for the, the, the network issue.

And another one, it's, you know, like 5G, now people do 6G, and I heard 7G somewhere, anyway. But for this kind of telecommunication studies, that means you need to use this kind of, it's not a line-based, right? It's a signal-based. If this can, whether this kind of signal can be captured by some other devices and store your data, right? So you need to use your own way to protect this kind of data in signals. So that's for the telecommunication part. So, okay, that's a little bit technical, but I will not talk about more with this kind of details, but general understanding is based on the app, we provide a clear way to describe how the AI community contributes to the technical side of the Responsible AI. Okay.

And then we also do a similar things. We try to link, uh, the AI's principles, uh, the Responsible AI's principles with several core AI techniques. Actually, this is quite similar with, if you still remember, a similar map for the SEO concerns here. So then I will skip off, skip off that one. But the results are quite similar. You could see the machine learning part is the AI one. This one, if you try to link all those kind of ties, actually they link with all those principles. Yeah. So that's to say, machine learning, the core of AI, so is standing in the core of the ethical concern or the responsible principles of this kind of the regulations of governance, these things. Okay.

Okay. So that's for the, um, mapping things. And then we also do similar things. But here it's slightly different because for the 18,000 articles, they generally based on the principles, based on the AI techniques. But here we don't focus on the journal articles or the conference articles, uh, published by the AI community, but we're looking to anyone who mentioned Responsible AI, this specific term, in their publications. Then we got a quite similar study with the AI ethics ones. The English-speaking countries, the leading universities or the leading research institutions were focusing on these kind of studies. Okay. But not that purely English-speaking. We still find like Sweden from Europe, the Italian, Netherlands, these kind of things. But also some top universities in Australia, actually. Yeah, Australia is quite a spearhead in Responsible AI. The CSIRO is somehow like our central, how that the central academy of science. So something like that one, but but not exactly, but it's a research institution. So they have a strong tie with the Australian federal government. So they sometimes or they usually, uh, could get the chance to represent for or on behalf of the Australian governments to attend various kind of conferences in, like the U in, uh, the UN or in the European Union for AI governance or Responsible AI, but some nowadays they call them Safe AI, but whatever. So that's why we, we, we, yeah, they provide several, like the opinion papers or comments to talk about their understandings about Responsible AI. Okay.

I think another interesting thing is here. This one, we try to track the relationships among those kind of, uh, the, the AI techniques, particularly Responsible AI related techniques, uh, since 2015. Why I do these kind of things? It's the starting point for deep learning is 2011, and the first paper in Nature, actually, the first paper mentioned AI ethics is a Nature paper. This one was published in 2015. So we studying our study from 2015. Actually, at that time, not that one mentioned ethics, but generally is a privacy related studies. Oh, that's what I say related to the password, the mimash, if you remember something I mentioned, the encryption. So that one. So they highlight how we store the password in a, you know, um, in a, how to say, comprehensive way or or in a confidential way or whatever these kind of things. But yeah, you can still see some cloud computing things, some other like trust was related to different kind of machine learning algorithms, and how can we do some empirical study with the information, as this kind of various kind of discussion on the AI ethics, and also like policy related studies. Anyway, so this map provides a sort of a landscape to track the evolution of those kind of Responsible AI related concepts, particularly techniques, from 2015 to 2025. Okay.