Imagine you are driving down a city street. You go around a bend and suddenly see something in the middle of the street in front of you. What should i do?
Of course, the answer depends on what that “something” is. A torn paper bag, a lost shoe or a tumbleweed? You can drive right over it without a second thought, but you will definitely be driving around a bunch of broken glass. You will likely stop for a dog standing in the street but stepping straight into a flock of pigeons knowing the birds will fly out of the way. You can plow straight through a pile of snow but drive around a carefully constructed snowman. In short, you quickly determine the actions that best fit the situation – what people call “common sense”.
Human drivers aren’t the only ones who need common sense. The lack of artificial intelligence (AI) will likely be the main obstacle to the widespread use of fully autonomous cars. Even today’s best self-driving cars are challenged by the problem of the object on the road. When these vehicles perceive “obstacles” that no human would ever stop for, they can unexpectedly apply the brakes and surprise other drivers. The rear end of human drivers is the most common self-driving car accident.
Cars need common sense to be trustworthy
The challenges for autonomous vehicles are unlikely to be solved by giving cars more training data or explicit rules for unusual situations. To be trustworthy, these cars need common sense: a broad knowledge of the world and the ability to adapt that knowledge in novel circumstances. While today’s AI systems have made impressive advances in areas from image recognition to language processing, the lack of a solid foundation of common sense makes them prone to unpredictable and inhuman mistakes.
Common sense is diverse, but one essential aspect is the mostly tacit “core knowledge” that people share – knowledge we are born with or which we learn when we live in the world. This includes a comprehensive knowledge of the properties of objects, animals, other people and society in general, as well as the ability to apply this knowledge flexibly in new situations. For example, you can predict that while a bunch of birds on the road won’t fly away when you get close, they will likely flock of birds. For example, if you see a ball bouncing in front of your car, you know that it may be followed by a child or dog to get it back. From this perspective, the term “common sense” seems to capture exactly what current AI cannot: use general knowledge of the world to act outside of previous training or pre-programmed rules.
Human learning versus machine learning
Today’s most successful AI systems use deep neural networks. These are algorithms trained to recognize patterns based on statistics obtained from large collections of humanly marked Examples. This process is very different from how people learn. We seem to be born with innate knowledge of certain basic concepts that help us find our way to understanding – including the notions of discrete objects and events, the three-dimensional nature of space, and the idea of causality itself. Humans seem Also being born with emerging concepts of sociality: Babies can recognize simple facial expressions, they have a clue of language and its role in communication, and rudimentary strategies for enticing adults to communicate. Such knowledge is so elementary and immediate that we are not even aware that we have it, or that it forms the basis of all future learning. A great lesson from decades of AI research is how difficult it is to teach machines such concepts.
In addition to their innate knowledge, children also show innate aspirations to actively explore the world, discover the causes and effects of events, make predictions, and induce adults to teach them what they want to know. Concept formation is closely related to children developing motor skills and awareness of their own bodies. For example, it seems that babies are starting to wonder why other people are reaching for objects at the same time so that they can reach for them themselves. While today’s cutting-edge machine learning systems start out as empty slates and act as passive, disembodied learners of statistical patterns; In contrast, common sense in babies grows through innate knowledge combined with learning that is embodied, social, active, and geared towards creating and testing theories of the world.
T.The history of the implantation of common sense in AI systems has mainly focused on cataloging human knowledge: manual programming, crowdsourcing or web mining commonsense assertions or computational representations of stereotypical situations. All of these attempts, however, face one major, potentially fatal, obstacle: Much (possibly most) of our core intuitive knowledge is unwritten, unspoken, and not even in our consciousness.
[Read: How do you build a pet-friendly gadget? We asked experts and animal owners]
The United States Agency for Advanced Defense Research Projects (DARPA), a major funder of AI research, recently launched a four-year program on Fundamentals of Common Sense that takes a different approach. It challenges researchers to develop an AI system that learns from “experience” in order to achieve the cognitive abilities of an 18-month-old baby. It may seem strange that matching a baby is seen as a major challenge for AI, but this reflects the gap between AI’s success in specific, narrow areas and more general, more robust intelligence.
According to developmental psychologists, core knowledge in infants develops along with a predictable timescale. For example, by the age of around two to five months, babies have knowledge of “object persistence”: if an object is blocked by another object, the first object is still there, although the baby cannot see it. At this point, babies also show awareness that when objects collide, they do not pass each other, but their movement changes. They also know that “agents” – entities with intent such as humans or animals – can alter the movement of objects. Between nine and 15 months, infants have a basic “theory of the mind”: they understand what another person can or cannot see, and after 18 months they can tell when another person shows the need for help.
Since babies under 18 months old cannot tell us what they are thinking, some cognitive milestones need to be derived indirectly. This usually involves experiments testing “violation of expectation”. Here a baby is observing one of two staged scenarios, only one of which meets the expectations of common sense. The theory is that a baby will look longer at the scenario that violates their expectations, and in fact babies tested this way will look longer when the scenario doesn’t make sense.
In the DARPA Fundamentals of Common Sense Challenge, each research team is tasked with developing a computer program – a simulated “Commonsense Agent” – that learns from videos or virtual reality. DARPA plans to evaluate these drugs by conducting experiments similar to those performed on infants and measuring the drugs’ “violation of expectation signals”.
This won’t be the first time AI systems have been assessed using tests designed to measure human intelligence. In 2015, one group showed that an AI system can match a four-year-old’s performance on an IQ test, leading the BBC to report that “AI had a four-year-old’s IQ”. More recently, Stanford University researchers created a reading test that formed the basis of the New York Post’s reporting that AI systems beat people for reading comprehension. However, these claims are misleading. Unlike people who do well on the same test, each of these AI systems was specifically trained in a narrow range and did not have any of the general skills that the test was designed to measure. New York University computer scientist Ernest Davis warned, “It is easy for the public to conclude that because an AI program can pass a test, it has the intelligence of a person who passes the same test.”
I think it is possible – even likely – that something similar will happen to DARPA’s initiative. It could create an AI program specifically trained to pass the DARPA tests for cognitive milestones, but not possessing the general intelligence that produces those milestones in humans. I suspect there is no shortcut to common sense, whether you’re using an encyclopedia, training videos, or virtual environments. To develop an understanding of the world, an agent needs the right kind of innate knowledge, the right kind of learning architecture, and the ability to actively grow up in the world. Not only should you experience physical reality, but also all social and emotional aspects of human intelligence that cannot really be separated from our “cognitive” abilities.
Although we have made remarkable advances, the machine intelligence of our present age remains cramped and unreliable. To create a more general and trustworthy AI, we may need to take a radical step backwards: we need to design our machines to learn more like babies, rather than specifically training them to succeed against specific benchmarks. After all, parents don’t directly train their children to receive signals that violate expectations. How infants behave in psychological experiments is simply a side effect of their general intelligence. If we can figure out how to make our machines learn like children, perhaps after a few years of curious physical and social learning, these young “commonsense agents” will eventually become teenagers – those who are sensible enough to be with the car key to be entrusted.
Published in collaboration with the Santa Fe Institute, a strategic partner of Aeon.
This article was originally published on Aeon by Melanie Mitchell and republished under Creative Commons.
Do electric vehicles excite your electrons? Do e-bikes make your wheels spin? Are you being charged by self-driving cars?
Then you need the weekly SHIFT newsletter in your life. Click here to login.
Published on March 11, 2021 – 09:19 UTC
Comments are closed.