Join us to recap the panel discussion with Yoshua Bengio(Turing Award Winner), Yolanda Gil(AAAI President) and Kevin Knight(ACL President 2011)! Dr. Guobin Wu, Director of DiDi Research Outreach will lead the discussion on the current state of artificial intelligence and the vision for the next generation of AI. What are the challenges and opportunities facing scientists, researchers and technologists…. 

 （From left to right: Guobin Wu, Yoshua Bengio, Yolanda Gil, Kevin Knight）

The Future of AI: Challenges + Opportunities  

Guobin Wu: Good evening, I am Guobin Wu. It's my great honor to host this panel to discuss The Future of AI: Challenges + Opportunities. First, please join me in welcoming our three panelists. They are:

Yoshua Bengio, a Canadian computer scientist, most noted for his work on artificial neural networks and deep learning. He was a co-recipient of the 2018 ACM A.M. Turing Award for his work in deep learning. He is a professor at the Department of Computer Science and Operations Research at the Université de Montréal and scientific director of Mila – Quebec Artificial Intelligence Institute.

Yolanda Gil, a Spanish computer scientist specializing in knowledge discovery and knowledge-based systems capture at the University of Southern California (USC). She has served as chair of SIGAI, the Association for Computing Machinery (ACM) Special Interest Group (SIG) on Artificial Intelligence, and as the president of the Association for the Advancement of Artificial Intelligence (AAAI).

Kevin Knight, Chief Scientist for Natural Language Processing (NLP) at Didi Chuxing. He leads a DiDi lab in Los Angeles devoted to NLP research. He was previously Dean's Professor of Computer Science at the University of Southern California (USC). Dr. Knight's research interests include human-machine communication, machine translation, and language generation. He served as President of the Association for Computational Linguistics (ACL) in 2011.

Again, thanks for accepting our invitation to join this panel. AI is very hot topic. If you don’t know how hot it is, just check the price of a hotel room near Long Beach, because of the ICML conference. Another example: in China, even middle schools introduce AI classes. You three are well-known as AI scientists. I hope today we can learn your insight about the future of AI.

Let’s start from a warm-up question. Would you please introduce your main efforts and achievement in your previous research on AI?

Yoshua Bengio: In the last 20 years, I would say a lot of my efforts has been in the notion that representations are very important but that they can be learnt, whereas classical AI was designing representations, machine learning tries to learn the representations, so we created a conference in fact, dedicated to this, ICLR, International Conference on Representations, which is now the home of a lot of deep learning research. And this has, of course, been the heart of what deep learning is about: deep learning is really about learning multiple levels of representations, so as to gradually build more abstract representations. Now the interesting thing is, we are still very far from that goal, even though there’s been huge progress in machine learning. Thanks to learning, humans and even children are able to discover and represent abstractions, in a way that’s well beyond the best AI systems, even if we have systems that can beat the champions of the game of the Go, and shine in many other fields, they are still in many ways limited compared to children, or even cats and dogs.

Yolanda Gil: I come from a tradition in AI where we worried about studying problem-solving knowledge, about reasoning about goals and tasks, and I’ve been personally very focused on scientific domains and, on scientific discovery as a target for AI. It i’s a very old tradition in AI to work in science domains, and it’s very exciting to see science advance because of AI techniques. I just came from a meeting with a lot of scientists that study climate, hydrology, and all kinds areas of geo-sciences. And what they have is tremendously fragmented pieces of data, pieces of knowledge, models of very specific processes, and integrating all those pieces is necessary in order to solve the problems that they are facing, such as, how is the health of certain ecosystem, how is the health of the planet do Earth processes work. Or in my own research, we are looking at regional scale, causality that relates climate to water availability and to human migration or to food security, and all of those problems phenomena have so many pieces across so many data sources and disciplines that they are a little bit beyond the human individual researcher scale. So I see AI as a technology that will really help scientists to do more integrative research that helps them so they can understand very complex phenomenon. So that is the case in environmental sciences and geo-sciences. That is also the case for example in neuroscience, we are trying to where understanding the brain and intelligence requires integrating very diverse pieces of knowledge. So I think AI has made a lot of strides and still has a lot of opportunities in moving propelling forward all aspects of scientific discovery.

Kevin Knight: I’ve been working twenty-five years in natural language processing, I started in natural language generation, trying to get machines to communicate thoughts and ideas to us using English or Chinese. I worked on one application that was on machine translation, trying to get machines that can translate between languages accurately. Or another of the applications, we were doing for example poetry generation system, which you might think it’s just for fun, but we found a lot of practical applications for poetry. For example, giving a password that’s easy to remember, because it’s in a poetic form. Or I recently worked on a Hollywood movie where we placed poetry throughout the movie, I think it’s the only movie just came out with texts that was written by a recurrent neural network intersected with a finite-state transducer. The Hollywood director doesn’t know what that means, but he’s totally satisfied with the output. We worked also to decipher historical manuscripts to help out our colleagues in the humanities to understand history better. That’s a number of things. Now I’m very excited about the work that we are doing in DiDi because there’s a lot of natural language data. If you think about passengers and drivers, customer service, autonomous vehicles themselves, all these things need to communicate with each other using natural language processing. If you leave your wallet in the car, you would like to talk to an automatic assistant that can help you get it back. And we want to talk English or Chinese if that’s your language. So lots of challenges ahead, lots more to do.

Guobin Wu: Recently, there are a lot of debates on Auto ML vs. Explainable ML. Some researchers think explainable ML will be the destination to solve everything, but some think Auto ML is good enough for specific scenarios like speech and image recognition. How do you see their debate?

Yoshua Bengio: Let me say a few words about explainable AI. If we look at human cognition, psychologists have tried to dissect a little bit the different kinds of tasks that we are able to do. There is one division which I like very much that I think helps understanding the limitation of current AI. And the question with explainable AI is the distinction between system 1 tasks and system 2 tasks. System 1 are things we can do very fast, they are very intuitive, we do them unconsciously, meaning we have no access to how the computation is being formed. For example, a lot of perception is like this. When you solve a task in half a second without having to think about it, that’s system 1. And system 2 tasks would be the kind that we can explain, that we do sequentially, slowly, that use explicit reasoning, and of course they are conscious and communicable. And I think that’s interesting because it’s saying that even humans have limits on how much they can explain their own decisions. Sometimes human experts will take a decision and then they will have to write a report, or answer questions like why did you do it this way? They will come up with a story. But it’s usually incomplete. Or about some high level aspects which might be very meaningful, but it wouldn’t be enough for computers using that high level knowledge to actually take those decisions. That’s one of the reasons why we are using machine learning, because makes it possible for computers to acquire the knowledge that we are not able to communicate to them explicitly through a language or some proxies like computer languages. The knowledge is provided through data and learning about it. So coming back to explainable AI, I think we’ll have to accept that all the tasks on which we will be using machines, the amount of explanation we are going to get out of that is going to be limited, it is going to be superficial, just like humans. If we were able to still make AI systems that could provide explanations as well as humans, that would already be great, but we are not there yet. So it is important to explore that kind of research. I’m interested in that interface between low level perception, the intuitive kind of computation, kind of what deep learning does, and the kind of explicit conscious linguistically expressible knowledge that can assist system 2 cognition. So we need to marry these things. Personally I’m trying to do it while staying in the framework of deep learning, just like your brain is doing it, with neurons, so there’s got be a way to do that. There’s been a number of proposals over the last decades but I think it’s still something in the forefront of research, and probably long-term. Short-term I think in terms of explainable AI, there’s a lot of little things that people are doing to take current deep learning systems, and extract information that can be useful to provide explanations, partial explanations, like which parts of the input was most responsible for the answer, this is something we know how to do. So there are very pragmatic things that can be done to deliver partial answers, it’s not a complete story, but can still be useful for users to explain behaviours of trained systems.

Yolanda Gil: I’m going to make some comments from the point of view of someone working on science domains, and the importance of automatic machine learning and also explanation in these domains. So for a long time, there is a long tradition in AI on, analyzing scientific data and coming up with learning novel patterns, sometimes discovering new laws that account for the data and so on. And one of the things that always comes to the forefront is that having an AI assistant that suggests new discoveries in a vacuum, and that is completely unaware of the world of science and what is already known in science, is not very helpful. When a scientist sees that a system has discovered a pattern or a particular law, they want to know, first of all, is it a new pattern? And second of all, how does it relate to all kinds of patterns or laws, or pieces of knowledge that we already have? And so I think aAs we move forward, our scientists will be much more willing and curious about using AI systems, if these AI systems are much more cognizant about what is known in that science. What is that gene? Why is it there? What is this storm? What is happening when the storm takes place? Why am I collecting this kind of data? So I think as AI systems become more knowledgeable about the world, they have more context about the data that they use, the environment where they operate, they will be much more helpful because they can explain their new discoveries in the context of what is already known. So that’s what I see in science. The explanations that they want are not about, there’s hardly an explanation that how this system works internally to reaches a conclusion, they are willing to trust that, you know, it is probably a smart system that has seen a lot of data and a lot of information reaches reasonable conclusion. What they want is an explanation along the lines of “I found this pattern and it’s significant because there’s all this knowledge that gives meaning and complements that pattern and shows why this discovery. It’s going to make a big and important change in your science.” So the explanation I always require to be in the context of what was learned. Beyond science, I believe that this sense of explanation as external context rather than internal workings of an AI system is actually most important. I have to say, at first without an explanation, that take for example recommender systems, I think that unless they take into account my personal preferences or my personal contexts, they are kind of lacking. So I don’t mind when a certain mapping system recommends that a specific path to me. I don’t care how it’s decided on that path, what. I care about is that it tells explains to me something like “I realize that you always prefer to take this street because it has beautiful trees and flowers.” So I want the explanations of how a solution fits my personal context and my needs, rather than the system looking inwards and just explaining how we reached a conclusion that we are willing to trust, but I want an explanation that suits me.

Kevin Knight: Journalists are always asking the famous song writer Bob Dylan, how did you come up with the lyrics of that song? He really hates that question, because what can you say? The system we were building for ourselves, as builders of the system and trainers, I actually do want to know what’s going on there. In 2015, we started building these neural network based machine translation systems. We put them under which you might think of as an MRI scanner and could find neurons that track the length of the sentence that just read, that actually track a lot of grammatical features of the sentence just read, whether it’s an active sentence or passive sentence, how many noun phrases there were. There’s a lot of grammar learning going on in these things. But when you think about that network as a whole, there’s way more knowledge about language than in any grammar book and that’s really the key. So it’s really interesting you can see what’s going on inside these things. But I think it’s tough to come up with a holistic and simple way for the machine to explain what it’s doing. Machines in the future might actually do as much rationalization as explaining, coming up with a reason to explain why it did what it did, even if that’s not the actual actual reason. So it’s a tricky, super-hard problem.

Guobin Wu: Next question is for Prof. Bengio. The development of AI is creating new opportunities to improve the lives of people around the world, from business to healthcare to education. It is also raising new questions about the best way to build fairness, interpretability, privacy, and security into these systems. These questions are far from solved, and in fact are active areas of research and development. I know you drive a lot of efforts on AI for good, Responsible AI, etc. Would you please share your viewpoints about these?

Yoshua Bengio: Yeah, certainly. As we are seeing the signature of this partnership today, AI has come out of universities and it’s being deployed in the real world. This is very different from previous attempts of taking AI research into the world because it’s at an amazing scale, it’s profitable, it’s used by a lot of companies and the investment is incredible. This is great. But you will have to be careful because what we are doing and researchers will give us more and more advances, it’s developing a very powerful tool. Humans are tool-makers. We used to make simple tools, and now we are making these very very powerful tools. And tools can be used in socially good ways or socially bad ways. Especially humans have propensities sometimes to take advantage of their power and tools can enhance that power. And so you have a potential vicious circle where the more powerful tools end up in the hands of the more powerful people, and they become even more powerful. That’s not good for society, right? Because it increases inequality, political instability, could lead to misery for many, not to speak of AI used in the military, which could destabilize the planet in many ways, that eventually could be bad for everyone. So I think AI scientists today feel a responsibility to be present in the social discussion, about how are we going collectively to use those tools. That’s why I think the more senior researchers are even more responsible. It is critical that we be involved in those discussions and learning as well because for example I was just very ignorant of a lot of social questions about how our societies are organized, ethics, etc.. In the last few years, I’ve learnt a lot because I had to learn about society, I was a part of lots of these discussions. I’ve talked to the philosophers, social scientists, people in health care. So Mila, in particular, has chosen to make these social impact, ethical questions as part of our mission, precisely because we want to own these responsibilities. I think it’s not just a duty. It also makes you feel good. So when you act in a responsible way like this, it makes you feel good about your work. If I were to design something that ended up killing a lot of people, I would probably feel bad. So I think this is from an emotional perspective, it’s really an opportunity for researchers in AI to claim that role. Maybe in an analogy it’s what happened around the second world war, with a lot of physicists who decided to rise, come out of their labs and start to be part of the collective discussion about what uses of nuclear physics was acceptable and what was not acceptable. I think this is a similar situation but now it’s not physics, it’s AI.

Guobin Wu: Yeah, I agree with you. Now government, industry and academia are paying attention to this topic. In 2018, DiDi founded AI for Social Good Platform in mobility industry, collaborating with over 10 universities, research institutions and social organizations. We focus on safety, health, environment and accessibility. For example, we installed remote sensors outside of DiDi vehicles to detect air condition so that we can draw high quality air condition map for improving environment analysis. Also, we developed driver-care AI assistants. These are highly aligned with Mila’s vision, both parties believe that with the development of AI technology, how to ensure AI in the premise of fair responsibility can empower society is of great significance.

Next question, let me ask Yolanda. As a professor and AAAI president, how do you think new trend of education system on AI, like establishing new institute of AI or new school of AI in universities?

Yolanda Gil: That’s a very good question, and a very hard one to answer. I want to say something about AI for Social Good, if I may. AAAI has made a partnership with a Chinese company called Squirrel AI. And we announced a Joint Annual Award of a million dollars on AI for benefit of humanity. We did this in Beijing last month. So I’m very proud of that because I think that it will highlight to the world that the profound benefits that AI could have on society and for the long-term well-being of humanity. So I’m very very excited about that. It will be an annual award and we’ll be setting it up very soon. So with respect to education…

Yoshua Bengio: If I could add something on this?

Yolanda Gil: Sure.

Yoshua Bengio: Thanks. So I think if we don’t do this, both individually and collectively, think through the social and ethical questions, eventually there will be a backlash. People already have fears about AI, they might be amplified when they see some of the negative uses. So we have to work on the positive, we have to make sure that society adopts the right social norms to make sure AI is used in a positive way for everyone. Otherwise, we are all going to lose.

Yolanda Gil: So l agreed. Let me talk about AI education then, I’ll be brief by making a comment based on analogy. So I think education for AI is going to become a lot more the way we look at education in the medical field. So in the medical field we have basic research in medicine, in biology, in physiology, you have a lot of basic research all of it is to understand the basic functioning of human body, of disease and so on. Similarly, there’s basic research in AI on different aspects of intelligence that’s very important and necessary and that continues to happen. But I also see that in the medical field you have hospitals that train doctors to do clinical and very practical applications of the science core research that’s being done in the lab. There are interventions, which are treatments. There are also a lot of people trained to support those doctors, there are nurses, there are technicians, there are social workers, there’s an entire ecosystem around how the hospital functions. We need to think about the future of the world as one where every single discipline, every single organization, every single institution is going to be completely reinvented through AI. And when we do that, it’s not going to be just the basic researcher, just the core ideas, but there will be an entire ecosystem of AI engineers, AI technicians, social workers, people who understand how processes have changed, and the role of humans in those processes, educators that train people to be up to speed with all the changes that AI will introduce continuously in any organization, and so on. So I thinkMy belief is that we need to think of AI education as an entire ecosystem of things professions that will arise over time. I give this analogy with the medical field because I think all of us will understand the diversity of jobs and responsibilities and positions. There are a lot of people in the medical area working on ethics, from the point of view of fundamental questions, from the point of view of bed side care, patient physicians, etc. So I see very similar things in terms of AI. I think universities need to start thinking about that kind of setting in order to design appropriate curricula for all those new to educate AI professions.

Yoshua Bengio: That’s difficult. I’ve been involved in the collaborations with people in healthcare from clinical to biological basic research. The scientific backgrounds are so far apart of computer science. It’s not easy. So, the education part is going to be a serious challenge. But recently, Montreal had an AI school for healthcare, a series of courses, they’re trying. It is not easy, but these kind of efforts on a much larger scale will be needed.

Guobin Wu: So, I do agree with Yolanda that the ecosystem is very important. In transportation section, DiDi has been contributing to the education as well. For example, DiDi GAIA open datasets have been supporting various educational endeavors around the world. A large number of students download the data for their graduation projects; over 10% among all GAIA applicants are teachers who use the datasets in teaching materials; and another 10% apply the data in STEM-related competitions. Moreover, DiDi scientists and researchers not only frequently enter university classes to give lectures, but we are also starting to develop high-quality online open courses to enable students from everywhere to learn about AI in transportation.

Third question is for Kevin. Talk about the roles among industry and academia in AI research. In terms of papers in ICML, academia contributes 77% and industry contributes 23%. From your perspective, what’s key for accelerating collaboration on AI research among industry and academia?

Kevin Knight: Let me first mention something about AI and ethics. Ten years ago, when we tried to translate for instance French into English, I remember a very famous translation, “Amy Winehouse was found dead in his apartment.” Doesn’t make sense, should be “her apartment”. This is system bias, because there is more “his” than “her” in the training material. And how the heck is it supposed to figure out that Amy is a female name, so it should say “Amy Winehouse was found dead in her apartment.” If you put it in Google Translate today, it comes out great. So, the smarter we make our system, the less sexist or racist it’s gonna be.

The tie-up between academia, industry and the government, on the ground, things like this happen: the Swedish national science foundation funded a student of mine to work on unsupervised translation, and he invented all kinds of stuff on that, then went off to industry to Google and led the team did Smart Reply in your Gmail system. So that’s one example. Someone else is a professor at John Hopkins who founded a spin-off company, and we had someone else do a spin-off company from USC in Natural Language Processing, that was quite successful, something hard to do. So on the ground, things like that happen every day, and we just have to make it work. Of course, tie-ups with more funding, government help, and academia creating new AI schools, all are critical.

Guobin Wu: Yeah, sure. More and more programs are working between industry and academia. So far, DiDi has established partnerships with many domestic and global academic Institutes, including Stanford University and the University of Michigan. DiDi also launched GAIA Open Dataset, providing desensitization data from the real world to the academic community, which will help scientists and researchers conduct basic and forward-looking studies in the field of transportation: 2800+ application from 660+ universities and research institutes from 30 countries.

Ok, now it’s the last question, and it’s for all of you. What is the next big thing for AI? In 2019, what is the biggest impression of AI from your perspective?

Yoshua Bengio: The whole idea of research is exploratory. Different researchers will have different answers to your questions. And it’s good because we don’t know which part is going to work better. So we should encourage different researchers, especially young researchers to explore and not just follow old people like me. That being said, I have my own opinions. Machine learning is at a transition phase, from the old pattern recognition type that it was very good at in the 90s, and it’s been really good in recent years, in many vision applications for example, to systems that will understand the world better. So coming back to what you are talking about, we are very far from having systems that understand the world even at the level of child. There’s a lot that needs to happen before we reach the point where scientists will be able to exchange with the AI system which understand the context. Before we will get to the point where that system will understand the scientific literature that’s relevant, we need to build systems that understand very simple things a child also understands, like intuitive physics, intuitive psychology, and all these very basic things. We are not there yet. Part of that might be we need much more training data, more multitasking learning everything, but I don’t think that’s enough either. I think fundamentally new principles are going to be added to our current tool box in order to approach human-level AI. So as I mentioned, we are moving from these simple function approximation, pattern recognition systems, to systems building a model of the world, systems which learn from interaction with their environment, that can capture the abstract explanation, that can capture the causal relationships, that can discover the high level discussion of what’s going on. One example of this is in natural language. So in natural language, if I look at the mistakes made by current systems, they are a lot better than where they were before, one thing I’m very proud of is what we have done in machine translation. But when you look at the mistakes these machine systems made, you realize that they don’t understand the world around us. They are making very stupid mistakes that humans don’t make. There’s a lot about the world which is not expressed verbally, that is what I was talking about with system 1 and system 2. Even if we have huge quantities of translated sentences, it might not be enough to build systems that converse with humans and actually make sense as well as human would. I believe that for those natural language understanding systems to also understand the world. If you don’t have an understanding on how the world works, no amount of processing from pure texts is going to save you. So that’s one direction where I see AI moving where we integrate natural language understanding, with vision, with robotics, with reinforcement learning, with systems that can seek knowledge just like what scientists or children are doing. One of the things human children are doing is playing, doing experiments, in order to acquire knowledge, in a way that’s purposeful. That’s not something current machine learning systems do but humans are very good at. So lots of interesting things for the future.

Yolaonda Gil: So I just co-chaired an effort and created a 20-year roadmap for AI research. It’s a public document. It was done with, with input from dozens of people from the community. There areIt highlights three major research areas in it that are very exciting for future research, I think. One is integrated intelligence, how can we combine individual assistance or individual capabilities to create composite capabilities --– this is in fact, composite tasks are very challenging. We don’t have a science of integration in AI, even though there has been an l. Lots of work on cognitive science. But architecture is not enough to address these questions. The second area is meaningful interaction. So not just being able to have a conversation with AI assistant but having a conversation where there are high stakes, where long-term memory of that interaction is important, and so on. Communicating not just through language but also many other channels. And the third area is self-aware learning. Learning systems that are able to understand their own limitations based on the data they have seen, to understand their roles in the world based on their awareness of their capabilities, what they have seen doing different tasks, and being able to set their own goals. So those are three big areas for the community.

I’d like to bring up one more exciting area if I may. I’m very interested. There is a particular way which I think AI is going to have a profound effect, which is that I believe AI systems are going to make us all better humans better. They are going to make us better at everything. They are going to be in watching judges and saying “You know, right before lunch you always make very harsh decisions. You know, just saying so you may want to think again about this one.” They are going to watch us drive, and tell us, “You know, you’re driving a little fast, given how much little you slept last night.” Whatever it is, they are going to really call on us, when we are discriminating, when we are not being safe, when we are not being fair logical. So I really see societies are moving forward in a way where these AI systems are starting to be deployed, they are going to help us learn more useful things about ourselves, improve our health, help us become better people. So that’s something I see AI researchers are addressing. How can we improve ourselves? I see a lot of interesting potential in this and that area. That i’s my personal wish for the future of AI.

Kevin Knight: It’s way too hard to say something that’s more interesting and fascinating than that. So, I’ll just keep it short. Future of AI, it’s too hard to tell. Maybe robots, maybe dialogue systems not just using language, but also integrating software systems and hardware systems. But I think it’s important that all the researchers do what they wanna do and try different things that they want. That will really increase the chance we hit the next big thing in AI.

Guobin Wu: Due to time limit, we have to end today’s panel. Let’s give a big applause to thank our panelists for their insightful opinions for the future of AI. What’s next big thing for AI, let’s wait and see! Thank you!