by Martin Ford
This has been my life’s challenge for the past 30 years. I published a book on that in 2000, with the second edition in 2009, called Causality. I co-authored a gentler introduction in 2015. And this year, I co-authored The Book of Why, which is a general audience book explaining the challenge in down-to-earth terms, so that people can understand causality even without knowing equations. Equations of course help to condense things and to focus on things, but you don’t have to be a rocket scientist to read The Book of Why. You just have to follow the conceptual development of the basic ideas. In that book, I look at history from a causal lens perspective; I asked what conceptual breakthroughs made a difference in the way we think, rather than what experiments discovered one drug or another.
MARTIN FORD: I’ve been reading The Book of Why and I’m enjoying it. I think one of the main outcomes of your work is that causal models are now very important in the social and natural sciences. In fact, I just saw an article the other day, written by a quantum physicist who used causal models to prove something in quantum mechanics. So clearly your work has had a big impact in those areas.
JUDEA PEARL: I read that article. In fact, I put it on my next-to-read list because I couldn’t quite understand the phenomena that they were so excited about.
MARTIN FORD: One of the main points I took away from The Book of Why is that, while natural and social scientists have really begun to use the tools of causation, you feel that the field of AI is lagging behind. You think AI researchers will have to start focusing on causation in order for the field to progress.
JUDEA PEARL: Correct. Causal modeling is not at the forefront of the current work in machine learning. Machine learning today is dominated by statisticians and the belief that you can learn everything from data. This data-centric philosophy is limited.
I call it curve fitting. It might sound derogatory, but I don’t mean it in a derogatory way. I mean it in a descriptive sense that what people are doing in deep learning and neural networks is fitting very sophisticated functions to a bunch of points. These functions are very sophisticated, they have thousands of hills and valleys, they’re intricate, and you cannot predict them in advance. But they’re still just a matter of fitting functions to a cloud of points.
This philosophy has clear theoretical limitations, and I’m not talking about opinion, I’m talking about theoretical limitations. You cannot do counterfactuals, and you cannot think about actions that you’ve never seen before. I describe it in terms of three cognitive levels: seeing, intervening, and imagining. Imagining is the top level, and that level requires counterfactual reasoning: how would the world look like had I done things differently? For example, what would the world look like had Oswald not killed Kennedy, or had Hillary won the election? We think about those things and can communicate with those kinds of imaginary scenarios, and we are quite comfortable to engage in this “let’s pretend” game.
The reason why we need this capability is to build new models of the world. Imagining a world that does not exist gives us the ability to come up with new theories, new inventions, and also to repair our old actions so as to assume responsibility, regret, and free will. All of this comes as part of our ability to generate worlds that do not exist but could exist, but still generate them widely, not wildly. We have rules for generating plausible counterfactuals that are not whimsical. They have their own inner structure, and once we understand this logic, we can build machines that imagine things, that assume responsibility for their actions, and understand ethics and compassion.
I’m not a futurist and I try not to talk about things that I don’t understand, but I did some thinking, and I believe I understand how important counterfactuals are in all these cognitive tasks that people dream of which eventually will be implemented on a computer. I have a few basic sketches of how we can program free will, ethics, morality, and responsibility into machines, but these are in the realm of sketches. The basic thing is that we know today what it takes to interpret counterfactuals and understand cause and effect.
These are the mini-steps toward general AI, but there’s a lot we can learn from these steps, and that’s what I’m trying to get the machine learning community to understand. I want them to understand that deep learning is a mini-step toward general AI. We need to learn what we can from the way theoretical barriers were circumvented in causal reasoning, so that we can circumvent them in general AI.
MARTIN FORD: So, you’re saying that deep learning is limited to analyzing data and that causation can never be derived from data alone. Since people are able to do causal reasoning, the human mind must have some built-in machinery that allows us to create causal models. It’s not just about learning from data.
JUDEA PEARL: To create is one thing, but even if somebody creates it for us, our parents, our peers, our culture, we need to have the machinery to utilize it.
MARTIN FORD: Right. It sounds like a causal diagram, or a causal model is really just a hypothesis. Two people might have different causal models, and somewhere in our brain is some kind of machinery that allows us to continuously create these causal models internally, and that’s what allows us to reason based on data.
JUDEA PEARL: We need to create them, to modify them, and to perturb them when the need arises. We used to believe that malaria is caused by bad air, now we don’t. Now we believe it’s caused by a mosquito called Anopheles. It makes a difference because if it is bad air, I will carry a breathing mask the next time I go to the swamp; and if it’s an Anopheles mosquito, I’ll carry a mosquito net. These competing theories make a big difference in how we act in the world. The way that we get from one hypothesis to another was by trial and error; I call it playful manipulation.
This is how a child learns causal structure, by playful manipulation, and this is how a scientist learns causal structure—playful manipulation. But we have to have the abilities and the template to store what we learn from this playful manipulation so we can use it, test it, and change it. Without the ability to store it in a parsimonious encoding, in some template in our mind, we cannot utilize it, nor can we change it or play around with it. That is the first thing that we have to learn; we have to program computers to accommodate and manage that template.
MARTIN FORD: So, you think that some sort of built-in template or structure should be built into an AI system so it can create causal models? DeepMind uses reinforcement learning, which is based on practice or trial and error. Perhaps that would be a way of discovering causal relationships?
JUDEA PEARL: It comes into it, but reinforcement learning has limitations, too. You can only learn actions that have been seen before. You cannot extrapolate to actions that you haven’t seen, like raising taxes, increasing the minimum wage, or banning cigarettes. Cigarettes have never been banned before, yet we have machinery that allows us to stipulate, extrapolate, and imagine what could be the consequences of banning cigarettes.
MARTIN FORD: So, you believe that the capability to think causally is critical to achieving what you’d call strong AI or AGI, artificial general intelligence?
JUDEA PEARL: I have no doubt that it is essential. Whether it is sufficient, I’m not sure. However, causal reasoning doesn’t solve every problem of general AI. It doesn’t solve the object recognition problem, and it doesn’t solve the language understanding problem. We basically solved the cause-effect puzzle, and we can learn a lot from these solutions so that we can help the other tasks circumvent their obstacles.
MARTIN FORD: Do you think that strong AI or AGI is feasible? Is that something you think will happen someday?
JUDEA PEARL: I have no doubt that it is feasible. But what does it mean for me to say no doubt? It means that I am strongly convinced it can be done because I haven’t seen any theoretical impediment to strong AI.
MARTIN FORD: You said that way back around 1961, when you were at RCA, people were already thinking about this. What do you think of how things have progressed? Are you disappointed? What’s your assessment of progress in a
rtificial intelligence?
JUDEA PEARL: Things are progressing just fine. There were a few slowdowns, and there were a few hang-ups. The current machine learning concentration on deep learning and its non-transparent structures is such a hang-up. They need to liberate themselves from this data-centric philosophy. In general, the field has been progressing immensely, because of technology and because of the people that the field attracts. The smartest people in science.
MARTIN FORD: Most of the recent progress has been in deep learning. You seem somewhat critical of that. You’ve pointed out that it’s like curve fitting and it’s not transparent, but actually more of a black-box that just generates answers.
JUDEA PEARL: It’s curve fitting, correct, it’s harvesting low-hanging fruits.
MARTIN FORD: It’s still done amazing things.
JUDEA PEARL: It’s done amazing thing because we didn’t realize there are so many low-hanging fruits.
MARTIN FORD: Looking to the future, do you think that neural networks are going to be very important?
JUDEA PEARL: Neural networks and reinforcement learning will all be essential components when properly utilized in causal modeling.
MARTIN FORD: So, you think it might be a hybrid system that incorporates not just neural networks, but other ideas from other areas of AI?
JUDEA PEARL: Absolutely. Even today, people are building hybrid systems when you have sparse data. There’s a limit, however, to how much you can extrapolate or interpolate sparse data if you want to get cause-effect relationships. Even if you have infinite data, you can’t tell the difference between A causes B and B causes A.
MARTIN FORD: If someday we have strong AI, do you think that a machine could be conscious, and have some kind of inner experience like a human being?
JUDEA PEARL: Of course, every machine has an inner experience. A machine has to have a blueprint of some of its software; it could not have a total mapping of its software. That would violate Turing’s halting problem.
It’s feasible, however, to have a rough blueprint of some of its important connections and important modules. The machine would have to have some encoding of its abilities, of its beliefs, and of its goals and desires. That is doable. In some sense, a machine already has an inner self, and more so in the future. Having a blueprint of your environment, how you act on and react to the environment, and answering counterfactual questions amount to having an inner self. Thinking: What if I had done things differently? What if I wasn’t in love? All this involves manipulating your inner self.
MARTIN FORD: Do you think machines could have emotional experiences, that a future system might feel happy, or might suffer in some way?
JUDEA PEARL: That reminds me of The Emotion Machine, a book by Marvin Minsky. He talks about how easy it is to program emotion. You have chemicals floating in your body, and they have a purpose, of course. The chemical machine interferes with, and occasionally overrides the reasoning machine when urgencies develop. So, emotions are just a chemical priority-setting machine.
MARTIN FORD: I want to finish by asking you about some of the things that we should worry about as artificial intelligence progresses. Are there things we should be concerned about?
JUDEA PEARL: We have to worry about artificial intelligence. We have to understand what we build, and we have to understand that we are breeding a new species of intelligent animals.
At first, they are going to be domesticated, like our chickens and our dogs, but eventually, they will assume their own agency, and we have to be very cautious about this. I don’t know how to be cautious without suppressing science and scientific curiosity. It’s a difficult question, so I wouldn’t want to enter into a debate about how we regulate AI research. But we should absolutely be cautious about the possibility that we are creating a new species of super-animals, or in the best case, a species of useful, but exploitable, human beings that do not demand legal rights or minimum wage.
JUDEA PEARL was born in Tel Aviv and is a graduate of the Technion-Israel Institute of Technology. He came to the United States for postgraduate work in 1960, and the following year he received a master’s degree in electrical engineering from Newark College of Engineering, now New Jersey Institute of Technology. In 1965, he simultaneously received a master’s degree in physics from Rutgers University and a PhD from the Brooklyn Polytechnic Institute, now Polytechnic Institute of New York University. Until 1969, he held research positions at RCA David Sarnoff Research Laboratories in Princeton, New Jersey and Electronic Memories, Inc. Hawthorne, California.
Judea joined the faculty of UCLA in 1969, where he is currently a professor of computer science and statistics and director of the Cognitive Systems Laboratory. He is known internationally for his contributions to artificial intelligence, human reasoning, and philosophy of science. He is the author of more than 450 scientific papers and three landmark books: Heuristics (1984), Probabilistic Reasoning (1988), and Causality (2000; 2009).
A member of the National Academy of Sciences, the National Academy of Engineering and a founding Fellow of the American Association for Artificial Intelligence, Judea is the recipient of numerous scientific prizes, including three awarded in 2011: the Association for Computing Machinery A.M. Turing Award for his fundamental contributions to artificial intelligence through the development of a calculus for probabilistic and causal reasoning, the David E. Rumelhart Prize for Contributions to the Theoretical Foundations of Human Cognition, and the Harvey Prize in Science and Technology from Technion—Israel Institute of Technology. Other honors include the 2001 London School of Economics Lakatos Award in Philosophy of Science for the best book in the philosophy of science, the 2003 ACM Allen Newell Award for “seminal contributions that extend to philosophy, psychology, medicine, statistics, econometrics, epidemiology and social science,” and the 2008 Benjamin Franklin Medal for Computer and Cognitive Science from the Franklin Institute.
Chapter 17. JEFFREY DEAN
We’re all working together on trying to build really intelligent, flexible AI systems. We want those systems to be able to come into a new problem and use pieces of knowledge that they’ve developed from solving many other problems to all of a sudden be able to solve that new problem in a flexible way, which is essentially one of the hallmarks of human intelligence. The question is, how can we build that capability into computer systems?
GOOGLE SENIOR FELLOW, HEAD OF AI AND GOOGLE BRAIN
Jeff Dean joined Google in 1999, and has played a role in developing many of Google’s core systems in areas like search, advertising, news and language translation, as well as in the design of the company’s distributed computing architecture. In recent years, he has focused on AI and machine learning and worked on the development of TensorFlow, Google’s widely-used open source software for deep learning. He currently guides Google’s future path in AI as director of artificial intelligence and head of the Google Brain project.
MARTIN FORD: As the Director of AI at Google and head of Google Brain, what’s your vision for AI research at Google?
JEFF DEAN: Overall, I view our role as to advance the state of the art in machine learning, to try and build more intelligent systems by developing new machine learning algorithms and techniques, and to build software and hardware infrastructure that allows us to make faster progress on these approaches and allow other people to also apply these approaches to problems they care about. TensorFlow is a good example of that.
Google Brain is one of several different research teams that we have within the Google AI research team, and some of those other teams have slightly different focuses. For instance, there’s a large team focused on machine perception problems, and another team focused on natural language understanding. It’s not really hard boundaries here; interests overlap across the teams, and we collaborate quite heavily across many of these teams for many of the projects that we’re working on.
We do deep collaborations with the Google product teams sometimes. We’ve done collaborations in the past
with our search ranking team to try to apply deep learning to some of the problems in search ranking and retrieval. We’ve also done collaborations with both the Google Translate and Gmail team, as well as many other teams throughout Google. The fourth area is researching new and interesting emerging areas, where we know machine learning will be a significantly new and important piece of solving problems in that domain.
We have quite a lot of work, for example, in the use of AI and machine learning for healthcare, and also AI and machine learning for robotics. Those are two examples, but we’re also looking at earlier-stage things. We have 20 different areas where we think there’s a real key aspect of some of the problems in that area that machine learning, or our particular kind of expertise, could really help with. So, my role is basically to try to have us be as ambitious as possible in all these different kinds of projects, and also to push us in new and interesting directions for the company.
MARTIN FORD: I know that DeepMind is heavily focused on AGI. Does that mean that the other artificial intelligence research at Google is geared toward more narrow and practical applications?
JEFF DEAN: That’s correct that DeepMind is more focused on AGI, and I think they have a structured plan where they believe if they solve this and this and this, that may lead to AGI. That’s not to say that the rest of Google AI doesn’t think about it. A lot of researchers in the Google AI research organization are also focused on building new capabilities for generally intelligent systems, or AGI if you want to call it that. I would say that our path is a bit more organic. We do things that we know are important but that we can’t do yet, and once we solve those, then we figure out what is the next set of problems that we want to solve that will give us new capabilities.
