Artificial intelligence does not become smarter until it understands the concept of causal relationships. It is not even given to all people.
In just a few years, computers have learned to diagnose diseases, translate texts and decipher speech. They beat people in complex strategic games, create photorealistic images and end sentences in letters for a person. However, despite these advances, artificial intelligence has drawbacks.
Machine learning systems are easily fooled or confused by unfamiliar scenarios. The algorithm of an unmanned vehicle does not know how to react to things that a human driver can easily handle. AI, which was trained to perform a specific task (for example, to distinguish cats), must be trained to perform another task (to distinguish dogs) again.
All the shortcomings of AI exist because artificial intelligence does not understand cause-effect relationships. The system sees that some things are connected with others, but does not understand that some things directly cause others. As if you knew that if the sky was overcast, it was more likely to rain, but you did not understand that it was raining from the clouds.
“Understanding cause-effect relationships is an important aspect of what humanity calls common sense, and this is an area in which artificial intelligence systems are“ ignorant ”today,” says Elias Barainboim.
His statement can be trusted: as director of the new Causal Artificial Intelligence Lab at Columbia University, he is at the epicenter of this problem.
His idea is to enrich research in the field of artificial intelligence with knowledge from a relatively new science of cause and effect relationships – an area heavily shaped by Jude Perle, a Turing Prize winner.
According to Bareinboim and Pearl, the ability of AI to detect correlations, for example, the fact that the appearance of clouds makes rain more likely is the simplest level of understanding of cause-effect relationships. But it turned out to be enough for the boom of artificial intelligence technologies based on deep learning to happen over the past decade.
Given the large amount of data on known situations, using this method you can make completely reliable forecasts. A computer can calculate the likelihood that a patient with certain symptoms has a disease, because he knows how often thousands of people with the same symptoms get sick.
But more and more scientists agree that AI progress will slow down if computers do not better understand cause-effect relationships. If machines could understand that some things could cause others, they would not have to learn all the time again – they could just take what they learned in one area and apply it in another.
And if computers could use common sense, we could trust more various tasks and not worry that they would make mistakes.
Modern AI is very limited in understanding the outcome of an action. In a learning system reinforced by signals from an interaction environment, a technique that allowed computers to master chess and go, the system makes extensive use of trial and error to determine which moves will lead to victory.
But in the real world with imperfect conditions, this approach does not work, the computer can not even in general terms understand how to play other games.
An even higher level of causal thinking is the ability to reason why something happened, as well as ask the question “What if …”
For example, a patient dies during a clinical trial. Is it the fault of the experimental medicine or not? Average school exams are getting worse. What reforms will improve them? This kind of reasoning goes far beyond the modern capabilities of artificial intelligence.
The dream of endowing computers with causal thinking forced Bareinboim to move to the United States from Brazil in 2008, after receiving his master’s degree in computer science from the Federal University of Rio de Janeiro. He grabbed the opportunity to study with Judah Perle, a computer science and statistics specialist at the University of California, Los Angeles.
Pearl is 83 years old, and he is a giant in the science of causation. His career helps to understand why it is so difficult to create artificial intelligence that thinks in terms of cause and effect.
Even qualified scientists are inclined to accept the relationship as a cause-effect relationship or, conversely, not dare to recognize a cause-effect relationship even when it is obvious.
For example, in the 1950s, several eminent statisticians did not recognize that tobacco causes cancer. They argued that without an experiment in which people would accidentally be prescribed the roles of smokers and non-smokers, it cannot be ruled out that some kind of stress, or possibly a gene, leads to both smoking and lung cancer.
In the end, the fact that smoking causes cancer has been conclusively proven, but it shouldn’t have taken so long. Since then, Pearl and other statisticians have developed a mathematical approach to determining what facts will be required to justify a causal relationship.
The Pearl method shows that, given the incidence of smokers with lung cancer, the existence of an independent factor that causes both is extremely unlikely.
The Pearl method also helps determine when correlation cannot be used to justify a causal relationship. Bernhard Shelkopf, who studies the technology of developing causation in AI as the director of the German Max Planck Institute for Intelligent Systems, indicates that the birth rate in the country can be predicted if you know the population of storks in this country.
This can be done not because storks bring children or children to attract storks to the country, but probably because economic development leads to the appearance of more children and storks. According to Shelkopf, Pearl helped give statisticians and computer experts ways to solve such problems.
Pearl’s research also led to the development of causal Bayesian networks – software that processes large amounts of data to find the variables that are likely to have the greatest impact on other variables.
GNS Healthcare of Cambridge, Mass., Uses this method to advise researchers on experiments that look promising. GNS has been working with researchers on one project to study multiple myeloma, a type of blood cancer.
Scientists wanted to find out why some patients with this diagnosis live longer than others after stem cell transplantation. The program processed data with 30 thousand variables and pointed to some of them, which seemed to be the most likely candidates for a causal relationship.
Biostatistics and experts in the field of myeloma eventually settled on one thing: the level of a certain protein in the patient’s body. After that, the researchers were able to launch a narrowly focused clinical trial to determine whether patients with this protein in the body really benefit more from the treatment.
“It’s a lot faster than sitting and guessing in the lab,” says GNS co-founder Iya Khalil.
Nevertheless, the progress that Pearl and other scientists have made in the theory of cause and effect has not yet brought significant successes in deep learning methods that reveal interdependence but do not see a causal relationship. Bareynboim is working on the next step: to make computers more useful tools for studying cause-effect relationships in humans.
One of the systems he developed, which is now in beta, can help scientists determine if they have enough data to answer a causal question. Richard McElrett, an anthropologist at the Max Planck Institute for Evolutionary Anthropology, uses this program to investigate the causes of menopause in humans (this is the only primate in which it is present).
The hypothesis is that fertility reduction in older women benefited early human societies because women who put more effort into caring for grandchildren ultimately had more descendants.
But what evidence today can be of the claim that it is more beneficial for children to be surrounded by grandparents? Anthropologists cannot simply compare the educational and health achievements of children who lived with grandparents to those who grew up without them.
Factors that statisticians call “interfering” intervene: grandmothers are more likely to live with their grandchildren, who need help most.
The Bareinboim program can help McElrett discern in which studies about children raised with grandparents, interfering factors are less present, which means it will help answer his causal question.
Barainboim speaks quickly and often waves with both hands, as if trying to balance the two sides of a mental argument. I came to him at Columbia University in October, in the middle of the semester, but it seemed like he had just moved into his office: the walls are almost empty, there are very few books on the shelves, only a graceful Mac and a white board, all covered with equations and diagrams, reminiscent of cartoon detail about crazy professor.
He dismissed the question of the uninhabited office, saying that he was very busy talking about both sides of the revolution in science about cause and effect. Barainboim believes that his work provides an opportunity not only to teach the causal thinking of the machine, but also to improve it in people.
Making people think carefully about cause and effect relationships is not much easier than teaching machines to think causally, he says. Researchers in many fields, from molecular biology to public policy, are sometimes content with correlations based on a causal relationship.
For example, some studies suggest that drinking alcohol will kill a person ahead of time, while others say: moderate consumption is safe and even beneficial. Another study shows that people who drink a lot live longer than non-drinkers in general.
This phenomenon, known as the “reproducibility crisis,” is present not only in medicine and nutrition, but also in psychology and economics. “You yourself see how unreliable all these conclusions are,” says Bareinboim. “Every couple of years we get a new result.”
He claims that anyone who asks “What if …” – doctors conducting clinical trials, sociologists developing pilot programs, web publishers preparing A / B tests – should start not with data collection, but with using Pearl’s causal logic and software like what Bareboym invented.
He hopes that “scientist” programs will appear in the future: a person is developing a cause-and-effect question, and software, using the theory of cause-and-effect relationships and machine learning methods, helps a person to exclude experiments that will not answer this question. This would help scientists avoid wasting expensive resources.
Bareynboim spoke about this in the lobby of the Sloan School of Management at the Massachusetts Institute of Technology.
According to him, this is too limited an approach, because human scientists conducting an experiment can take into account only a few variables in their minds at a time. But the computer can see the interaction of hundreds or thousands of variables.
A computer scientist, whose code contains the basic principles of Pearl’s cause-and-effect calculation, is able to find out what can happen with a new set of variables, can suggest which experiments human researchers should spend their time on.
Perhaps some state practice that works only in Texas could be used in California, with more emphasis on several cause and effect factors. “Scientists will no longer“ experiment in the dark, ”says Bareinboim.
This finish line will probably come with the help of technologies that are just starting to develop. Joshua Benggio, a computer specialist from the University of Montreal, one of the 2018 Turing laureates, is trying to teach neural networks, the programs that underlie deep learning, the method of “meta-learning” – to determine the causes of events.
In order for a neural network to determine when a person is dancing, you need to show her many images of dancers. In order for her to determine when a person is running, she needs to show her many images of runners. The system will learn to distinguish runners from dancers, revealing the features of each group of images, for example, the position of the fingers and hands of a person.
But Benjio points out that fundamental knowledge about the world can be obtained by analyzing several data arrays and finding similar, or “invariant” things. Perhaps the neural network will be able to understand that the physical movements of the legs cause both running and dancing.
Maybe seeing these and many other examples that show people at a height of just a few centimeters from the ground, the machine will eventually understand something about gravity and how it restricts human movements.
Over time, with a sufficient amount of meta-study of variables that are identical in different data arrays, a computer can obtain causal knowledge that could be reused in many areas.
Pearl argues that AI cannot be truly “smart” until it can fully understand cause-effect relationships. Although only causal thinking is not enough for general-level artificial intelligence, it is necessary, he says, because it would allow the AI to reflect, that is, to conduct the process that underlies cognition.
“Questions of the category“ What if … ”are the structural elements of science, moral principles, free will, consciousness,” says Pearl.
Pearl refuses to make predictions about how long it will take for the computer to get full-fledged ability for causal thinking. “I’m not a futurologist,” he says.
Nevertheless, Pearl is convinced that the first step should be the development of machine learning tools that combine data with existing scientific knowledge: “Our brains contain a lot of information that we do not use at all.”