Artificial intelligence (AI) became a scientific field almost 60 years ago. Ever since then, researchers have tried to achieve human-level smarts or better.
Yet even with recent feats of computational genius — for example, Google DeepMind beating a human player in the game Go — AI scientists say they still have a long road ahead.
Tech Insider spoke with AI researchers, computer scientists, and roboticists around the world about what it is going to take to build a machine that’s able to think, work, and feel like a human.
Here are their lightly edited responses.
Bart Selman said computers need to learn how to understand the world like a human.
“The big obstacle, though it’s not an obstacle because I think it will just take time, is the computer has to learn more about the way we see the world.
“It’s very hard to understand the world from a human perspective. Intelligence relies on the way we view the world as humans, and the way we think about the world.
“Computers are just starting to be able to hear and starting to being able to see images. Those are tremendous improvements in the field in the last five years.
“We’re doing that by having computers read millions of texts and pages from the web, by hooking them up to cameras and moving them around human environments.”
Commentary from Bart Selman, a computer scientist at Cornell University.
This experience of the world will foster more intelligent AI, Peter Norvig says.
“AI needs to experience living in the world.
“We are very good at gathering data and developing algorithms to reason with that data. But that reasoning is only as good as the data, which, for the AI we have now, is one step removed from reality.
“Reasoning will be improved as we develop systems that continuously sense and interact with the world, as opposed to learning systems that passively observe information that others have chosen.”
Commentary from Peter Norvig, director of research at Google.
To do that, Yoshua Bengio says computers should be trained to learn like children.
“Right now, all of the impressive progress we’ve made is mostly due to supervised learning, where we take advantage of large quantities of data that have already been annotated by humans.
“This supervised learning thing is not how humans learn.
“Before two years of age, a child understands the visual world through experiencing it, moving their head and looking around.
“There’s no teacher that tells the child, ‘in the image that’s currently in your retina, there’s a cat, and furthermore it’s at this location’ and for each pixel of the image say ‘this is background and this is cat.’ Humans are able to learn just by observation and experience with the world.
“In comparison to human learning, AI researchers are not doing that great.”
Commentary from Yoshua Bengio, a computer scientist at University of Montreal.
Yann LeCun echoes the idea that computers need to be more like babies.
“The short answer is: we have no idea. That’s why it’s very difficult to make predictions as to when ‘human-level AI’ will come about.
“Right now, though, the main obstacle we face is how to get machines to learn in an unsupervised manner, like babies and animals do.”
Commentary from Yann LeCun, director of Facebook Artificial Intelligence Research.
That includes learning a kind of common sense, says Ernest Davis.
“The lack of common sense reasoning is a major obstacle. There’s large amounts of basic understanding of the world that we haven’t been able to get programs to do.
“To a very large extent, AI programs work by avoiding the problem, work by getting around the understanding of language or understanding what the text means and what it says about the world. There are limits to that.
“Before we can get fully intelligent programs, those problems are going to have to be overcome. But we don’t know how we ourselves understand the world, mostly, so these traits are incredibly difficult to emulate in a program.”
Commentary from Ernest Davis, computer scientist at New York University.
Murray Shanahan notes that computers also need to be more creative.
“The two biggest obstacles to human-level AI are endowing computers with common sense and endowing them with creativity.
“By endowing them with common sense, I mean giving computers the ability to understand the consequences of everyday actions — actions on physical things or social actions, things that you say or do with other people.
“By creativity, I don’t mean the kind of thing that Mozart or Einstein could do, but the kind of thing that every child is capable of when they play.
“That kind of creativity is just being able to come up with completely new sorts of behavior and to explore them in an effective way, which children are amazing at. That’s how they learn.
“We really don’t know quite how to endow computers with those two capabilities yet.”
Commentary from Murray Shanahan, a computer scientist at Imperial College.
Carlos Guestrin said computers need to be able to understand abstract concepts like humans do.
“Humans have developed what are called abstractions — we can think about, for example, cars generically without thinking about a specific type of car, but we can also quickly dig down to a specific car and to parts of cars. We can learn about those abstract notions and those specific notions very quickly through only a few examples.
“Today, it takes computers a lot of data, a lot of examples of cars, to learn what a car is. Then to generalize and represent different levels of abstractions or different levels of specificity is extremely difficult.
“That’s one big gap between what a human can do, even a child can do, and what a computer can do.”
Commentary from Carlos Guestrin, the CEO and cofounder of Dato, a company that builds artificially intelligent systems to analyze data.
The fundamental problem is that AI know can only do what they’re told to do, says Michael Littman.
Robot HRP-2 demonstrates use of a tap after washing a cup at Tokyo University
“Some of the great successes lately have been things like deep learning — methods that take lots and lots of data and then are able to mimic human judgments about that data. Things like object recognition — where you show the computer a picture and it can label that picture, ‘oh that’s a woman by the beach,’ that kind of thing.
“What’s missing from a lot of these systems at the moment is a notion of a will — a desire to do something in the world. It’s just doing what it’s told, which is to map inputs to outputs.
“There’s not a lot of room for creativity there. The kinds of problems we are asking these systems to do are not really on a path towards a sentient system.”
Commentary from Michael Littman, a computer scientist at Brown University.
Matthew Taylor says computers just can’t do as many things as humans can.
“Programs that think like humans are so far beyond where we are right now. I don’t think the field even knows the right direction to in go to achieve that goal, let alone what the big roadblocks are.
“It would be really nice if we had artificial general intelligence but that’s a long way away. Even winning Jeopardy, that’s still a very constrained situation.
“That computer didn’t need to know anything about how to walk or how to move, it didn’t need to know that water is wet, or that people eat food — all this basic stuff that we take for granted.
“Computers have a ways to go before they can even use that kind of knowledge.”
Commentary from Matthew Taylor, a computer scientist at Washington State University.
Toby Walsh says computers need to be more adaptable.
“We have some pretty good examples of superhuman performance, computers able to perform at levels which exceed those of humans.
“In the game of chess, answering questions on Jeopardy! — computers have clearly, demonstrably proved themselves able to perform at the level if not above the level of humans.”
“But the thing that humans are just so wonderful at is our adaptability, our ability to work in new situations. If I parachute you into a new situation, you will very quickly adapt and work in that situation.
“Computers are still driven and very focused on what they’ve been told to do — they’re very unadaptable at working.
“If you change the parameters, like “don’t play chess, play backgammon.” The Deep Blue chess-playing program is no good at backgammon. Getting the breadth of ability of humans is certainly going to be a big challenge.”
Commentary from Toby Walsh, a professor in AI at the National Information and Communications Technology Australia.
Shimon Whiteson said we need to figure out how to get robots to move quickly between tasks.
“If we are talking about building a system that has intelligence which is at a comparable level to that of a human being, one bottleneck is computational power. The hardware just needs to get better. But this is a very transient obstacle because computers are getting faster all the time.
“It will only be a matter of years before we have much more powerful computers that can do things we can’t even imagine today. When the computational power is here I think we have good algorithms for doing AI, good algorithms for learning.
“But there are some things we can’t do well. Humans, for example, are really good at generalizing in different situations.
“You learn one task and can very quickly apply what you’ve learned to a different task, even if the relationship between the tasks is not that obvious. This is something that computers are really bad at that we don’t have good algorithms for.”
Commentary from Shimon Whiteson, an associate professor at the Informatics Institute at the University of Amsterdam.
Thomas Dietterich says we need to build computers that can do different things.
“When people talk about ‘human-level AI’ they typically mean ‘human-breadth AI.’
“Adult humans can answer questions and solve problems across a wide range of activities — finance, sports, child rearing, collaboration, opening packages, planning trips, packing a car, and shopping for a vacuum cleaner.
“No AI system comes anywhere close to having this immense breadth of capabilities, particularly when it comes to combining vision, language, and physical manipulation.
“Additionally, we don’t even know how to represent the knowledge and information that is needed for all of these different tasks.
“Hardly any AI research groups are studying this question. Instead, they are focused on improving computer performance on narrow tasks.
“It is easier to make progress and measure success on narrow tasks, whereas it is difficult to develop useful measures of general intelligence. We hardly know where to begin.”
Commentary from Thomas Dietterich, the President of the Association for the Advancement of Artificial Intelligence.
To build a computer that can do all the different things a human can, Subbarao Kambhapati says two areas that have developed in isolation need to come together.
“There is great progress being made in understanding low level perception — being able to see the world, hear the world, touch the world, open doors — a very complicated, very technically challenging set of capabilities.
“We have also made progress at the top level — being able to reason about other people’s minds, and then trying to anticipate their actions.
“But these pieces have been done separately, and bridging them is going to be a very important challenge.
“For example, when humans see a surveillance video, we start making a story up in their minds, that this guy is trying to do the following thing and this other guy is trying to do the following thing.
“That ability requires two different sorts of processing — one is low-level vision, and the other is higher-level reasoning about people’s goals and intentions. They need to be combined.”
Commentary from Subbarao Kambhapati, computer scientist at Arizona State University.
Lynne Parker says we don’t have a single clue about how to build thinking, conscious machines.
“I don’t have an answer for the obstacles in our way to building human-like robots, because we just don’t know enough about how people reason. We haven’t figured out the fundamental principles, so I don’t think we know what the hurdles are.
“Maybe we need to find a way of representing and interconnecting facts, knowledge, and methods of reasoning — the whole structure of what intelligence looks like might be necessary to understand in order to achieve sentient reasoning in AI, but we don’t even know what that looks like in the human brain.
“Whether or not we can achieve that without mimicking the human brain, I don’t know. But there’s something fundamental that we’re not getting, in terms of how AI should be structured.”
Commentary from Lynne Parker, the division director for the Information and Intelligent Systems Division at the National Science Foundation.
Particularly because we don’t know how humans reason or create consciousness, according to Manuela Veloso.
“The obstacles are that we don’t know much about how humans reason.
“We do know that humans can make a lot of decisions that are connected to functions. They can cross roads, they can pick up objects, they can construct things.
“We have to understand how people do it, but we also have to define the problems in a way that we can find algorithms to do them.”
Commentary from Manuela Veloso, a computer scientist at Carnegie Mellon University.
Stuart Russell thinks we need to knock consciousness off its pedestal.
“I used to say that if you gave me a trillion dollars to build a sentient or conscious machine I would give it back. Because I could not honestly say I have any idea how it might work.
“Consciousness is a first person subjective experience. We don’t have any scientific theory of any kind that could lead us to a detailed map of it.
“Even if we could simulate someone’s brain in exquisite detail, there’s nothing in any scientific theory that I’m aware of that can tell us that the operation of that particular physical system would generate a conscious experience.
“We don’t have even the beginnings of a theory whose conclusion would be ‘such and such a system is conscious.'”
Commentary from Stuart Russell, a computer scientist at the University of California, Berkeley.
Yoky Matsuoka says that it’s not even clear that just adding more computational power will result in intelligence.
“One of the problems with AI is that it’s very much a black box approach.
“Deep learning is a good example. Because it’s such a black box, it’s difficult for humans to comprehend everything that’s going on inside the neural networks.
“To advance that, the only sort of knowledge we have is to say ‘let’s increase the number of neurons, number of connections, the computational power to increase the memory.’
“If we really build the number of neurons similar to the human brain and made all the right connections and started putting the same inputs in, are we really going to achieve human-level intelligence?
“We don’t know. That’s the problem.”
Commentary from Yoky Matsuoka, former vice president of technology at Nest, a Google-owned company that makes smart thermostats.
On the other hand, Geoffrey Hinton says consciousness is besides the point.
“The biggest obstacle is the idea that there is some mysterious essence called “consciousness” that is required to make things sentient.
“Consciousness is an old and very primitive attempt to explain what’s special about a very complicated computational system — the human brain — by appealing to some unobserved essence.
“The concept is no more useful than the concept of ‘oomph’ for explaining what makes cars go.
“It’s true that some cars have a lot more oomph than others, but that doesn’t explain anything about how they work.”
Commentary from Geoffrey Hinton, researcher at Google and computer scientist at the University of Toronto.
Samy Bengio says AI may not even go in the direction of an sentient, all-knowing robot.
“Something like ‘human intelligence,’ it’s not even clear AI research will evolve in that direction one day.
“Many pieces are still missing — in particular a better use and interaction of a long term memory and perception-based models. Models, like the Neural Turing Machine or Memory Networks, can achieve some of this, but it’s only the beginning.
“We also need to be able to learn more with unlabeled or less labeled data, which is still an open research problem.
“We need to work more on continuous learning — the idea that we don’t need to start training our models from scratch every time we have new data or algorithms to try. These are simply very difficult tasks that will certainly take a very long period of time to improve.”
Commentary from Samy Bengio, researcher at Google.
For now, we should focus improving what’s possible, Sabine Hauert says.
“The first challenge is actually understanding what sentient reasoning means. Nobody has a good definition of that. I personally believe that we’re very far from anything that is deemed to be sentient.
“The second obstacle is the way we communicate about AI. There’s a lot of hype and a lot of misconception around robotics and AI.
“A lot of it has to do with using words like ‘sentient reasoning’, which is very far from anything that’s there already. As a result of that, I think we’re not talking about the right things.
“Instead of talking about sentient reasoning, I think we should be spending time thinking about the technologies that are much closer to happening.
“Trying to understand what their benefits are, what their impact is to society, what the legal and safety questions are.”
Commentary from Sabine Hauert, a roboticist at Bristol University.
This article was originally published on Tech Insider. Read the original here.