blog posts

Can Computers Learn Common Sense?

Can Computers Learn Common Sense?

Humans Have A Gift That Computer Scientists Have Been Unable To Inject Into The World Of Computers To This Day. Can Common Sense Be Taught To Computers?

Common Sense, A few years ago, a computer scientist named Yajin Choi spoke at a scientific conference on artificial intelligence in New Orleans. He showed a picture of a local news section on a projector screen in his speech. Two newscasters were seen next to a prominent news headline in this image. The news was: stabbing a cheeseburger.

Choi said in his speech that a person could understand the generality of the news by looking at these two words. Has anyone stabbed a cheeseburger? Did one person use a cheeseburger to stab another? Probably not.

Has the cheeseburger attacked another cheeseburger with a knife? It seems very unlikely. This scenario is also definitely impossible. The only possible interpretation of this headline is that one person stabbed another person for arguing over a cheeseburger.

Computers fail to understand some of the issues that seem obvious to us

Choi went on to say that it is difficult for computers to solve this simple puzzle. Computer programs have one major weakness: the deprivation of common or common sense.

General understanding or logic is the realistic understanding that human beings have implicitly acquired throughout life and do not need a specific reason to prove it correct. Common sense, for example, says that jumping from the top of a building or touching a hot object can cause harm, so people avoid doing these things. It is the knowledge that computers lack, so they can not rule out the possibility of violence between cheeseburgers because it is irrational.

Artificial intelligence can be as skilled as or even more capable than human thinking in performing specific tasks, such as playing chess or diagnosing tumors. Still, there are countless unpredictable situations in the real world, and this is where artificial intelligence comes into its own.

 Researchers use the term “marginal” to describe these conditions, that is, situations outside the scope of regular or predictable conditions.

The human brain artificial intelligence

Human beings use their common sense to deal with problems in these situations. Still, the performance of AI, which often depends on pre-written and hard-and-fast rules or logical communications, can not cope with the challenges posed in these unusual situations.

All human beings have common sense or innate logic, an ability we may not always think about, But imagine a world in which no one behaves logically. The importance of this ability is emphasized more than ever.

Imagine for a second if transposed you into the karmic-driven world of Earl.

If you as a robot are deprived of logical thinking, you think to yourself how your body shape may suddenly change.

On the way home, you may see a broken fire hydrant on the side of the street, and the water is erupting from it and wetting part of the street; you can not tell if it is safe to drive under a fountain of water. Is it or not. Then you park your car outside a pharmacy; next to the pharmacy entrance, you see a man lying on the ground bleeding and asking for your help.

In this situation, are you allowed to quickly remove the dressing from the pharmacy and rush to the aid of the person without waiting in line at the pharmacy to pay for them, or you may encounter a news report about the stabbing of a burger when you get home and turn on the TV?

We all do it every day because life is full of these fringes. As you can see, living in the real world is exhausting for a robot without a general understanding of the world. Still, a human being, relying on a treasure trove of tacit knowledge and awareness, can correctly interpret these conditions and make the right decision.

Life in the real world is a pain in the ass for a robot

“A logical understanding of our lives is like a dark matter in astronomy because it is ubiquitous and shapes most of what humans do,” says Oren Etzioni, director of the Allen Institute for Artificial Intelligence in Seattle. However, it isn’t easy to describe.

The Allen Institute works with the United States Government on Advanced Defense Research Projects Agency (DARPA) to study and develop artificial intelligence systems. In fact, in 2019, the DARPA agency launched a $ 70 million four-year program called Machine Logic. Many complex problems in this field will solve if scientists succeed in teaching logic to their computer programs.

An article on the subject of artificial intelligence provided compelling examples of the inclusion of logic in computer programs; A general-understanding robot, looking at the surface of a table and seeing a piece of wood protruding from its surface, concludes that the amount of wood is probably part of the surface of the table itself, not a random selection of wood thrown on the table.

A language translation system can easily understand language ambiguities and punctuation.

A home cleaning robot learns that it should not throw the cat in the trash or fold it up and put it in a drawer. Such systems can only work in the real world if they have an inherent logic and a general understanding of the world, something we do not think is worth much.

In the 1990s, questions about artificial intelligence and the issue of computer security led Etzioni to study computer logic. In 1994, he co-authored an article on the formulation of the first law of robotics. The Laws of the World of Robots is an imaginary instruction written by Isaac Asimov in his science fiction novels. According to the first law, a robot should not harm a human, or a robot should not allow it; As a result of not interfering in a situation, a human being is harmed.

However, the problem Etzioni encountered in explaining these laws scientifically in the first place was that robots had no understanding of the concept of harm. Utilizing such knowledge requires a broad and moral sense of human needs, values, and priorities. Otherwise, mistakes by robots will be inevitable.

Robots have no understanding of the concept of harm

In 2003, for example, Nick Bostrum, a philosopher, wrote an article describing a situation in which an artificial intelligence system asked to maximize paper clips production in a factory.

After a while, the system realizes that humans may sometimes shut him down, resulting in reduced production, so he decides to get rid of humans to complete his mission.

Bostrom’s AI system does not smell of logic; it may tell itself that scrambled papers that do not use paper clips to sort them are a kind of damage, But trying to incorporate logical understanding into computer programs can present many challenges.

In recent years, computer scientists have categorized samples of so-called hostile data. In this data category, minor changes have been made compared to actual models to mislead artificial intelligence systems.

One study found that placing a few harmless small stickers on specific parts of a stop sign could cause computer image recognition systems to confuse the stop sign with a speed limit sign.

In another study, a slight change in the printing pattern of a three-dimensional model of a turtle caused an AI computer program to confuse it with a rifle.

If this AI program had a general understanding, he would not be so easily deceived and know that guns do not have a lock and four legs.

The human brain and the computer

Yajin Choi, who teaches at the University of Washington and works at the Allen Institute, says of the history of artificial intelligence that in the 1970s and 1980s, researchers thought it was close to programming logic and integrating it into computer programs. Still, later they concluded that “no, it seems too difficult,” so they went for more specific challenges such as image recognition and language translation; But today, the situation has changed.

Many artificial intelligence systems, such as self-driving car technology, may be present in the real world alongside humans shortly. It has made the need for artificial intelligence systems to have a general understanding of the world more than ever.

Teaching logic to computer programs also seems more accessible today than ever before.

Computers have improved human-independent learning, and researchers can now feed computer programs with more appropriate data. Artificial intelligence could soon cover more of the world.

How does rational understanding develop in humans? The short answer is that we are multidimensional learners. We experiment with some things and observe our work results, read books and listen to instructions, and sit silently watching and arguing with ourselves.

In the course of life, we ​​fall face down and look at the mistakes of others, But unlike humans, artificial intelligence systems lack these learning skills. They often want to go to the end, leaving all the routes aside.

In 1984, a computer scientist named Doug Lenat began developing a logical encyclopedia called Psych. This encyclopedia was based on axioms, general rules, and laws explaining how the world works. The researchers first chose to provide specific instructions to the programs.

For example, according to one of these rules, owning one thing means owning all its components. Another rule says that complex objects can damage soft objects. The third rule may state that human flesh is more delicate than metal.

By combining these rules, a logical conclusion can reach: If the bumper of your car hits a person’s foot, you will be responsible for the injury. Lenat explains this system:

Psychic uses instantaneous and complex network expressions to understand the situation instantly and argue about it.

Cycorp, the company that developed the Psychic program, has been around for decades, during which time hundreds of local people have introduced tens of millions of rules to the program. Cyclops considers its products confidential and does not provide much information about them. Still, Stephen DeAngelis, CEO of Antra Solutions, a consulting firm for manufacturing and retail companies, believes Scorpion’s software can be powerful.

He uses a culinary example to illustrate his point. He says that the psychic program has a great deal of knowledge about the characteristics of foods and the taste of fruits and vegetables and can rely on his logic to argue that although a tomato is a fruit, it should not use in fruit salads.

However, researchers say that developing psychics is a thing of the past. They say that it is impossible to include all the details of life in the real world by writing rules in a computer program.

Instead of handwriting instructions, they focus on another solution called machine learning, the same technology used in products such as Siri, Alexa, Google Translate, and other digital services, based on recognizing patterns from large amounts of data. Is. Instead of reading instruction manuals, machine learning systems analyze libraries in the sense of their digital world.

It is impossible to include all the details of life by writing rules in a computer program.

In 2020, the OpenAI Research Laboratory unveiled a machine learning algorithm called the GPT-3. This computer model can produce meaningful and human-like texts by studying texts on the World Wide Web and discovering language patterns.

Imitation of this system of human writing patterns is unique in some ways, but strange mistakes are made in some cases. This system sometimes produces weird writing; for example, in one case, the GPT-3 program compiled this text: To jump from Hawaii to seventeen, you need two rainbows. If this program had logic, it could understand that the rainbow is not a unit of time and the number seventeen is not a place.

Choi and his team planing to use language models such as the GPT-3 as a cornerstone to introduce logic to AI programs. In another series of studies, GPT-3 was asked to generate millions of IH logical expressions to describe a cause-and-effect relationship and express decisions, such as “Before Lindsay is offered a job, she must apply. Offer.”

Choi’s research team then asked another machine learning system to analyze a filtered set of expressions to answer the questions designed to fill in the blanks finally.

The evaluation of the answers provided by the artificial intelligence system to these questions shows that in 88% of the cases, the answers to the questions were utterly logical, which offers a significant improvement compared to the statistics of the GPT-3 system with 73%.

The Choi-led lab conducted a similar experiment on data in the form of short videos.

Choi and his colleagues first created a database of more than a million tagged clips and then used an artificial intelligence system to analyze it. Internet users were then asked, for a fee, to ask multiple-choice questions from short videos and clips that differed from those shown to the AI ​​system.

In addition to answering multiple-choice questions, the system had to explain why it answered—selected A frame from the Swingers movie in one of these questions, where a restaurant waiter places a pancake order on a table.

Three men are sitting around the table, and one of them is pointing at the other. In answer to the question “Why does person 4 refer to person 1?”

The system responded that “the pointer tells Person 3 that Person 1 ordered the pancakes.”

When asked the system to explain its response, it explained that “the person has placed three customer orders on the table and may not know which one belongs to the order.” This artificial intelligence system provided logical answers in 72% of cases. This figure is not much different from the response statistics provided by human users, which is 86 percent. Such systems perform amazingly well,

The world of computers and zeros and ones

These systems seem to have enough logic to understand everyday situations in terms of physicscause and effect, and even psychology. It can even say that the AI ​​program knows that people are eating pancakes in dining halls and that people may have given different orders, and that hand gestures can be a way of transmitting the information.

The performance of AI systems is impressive in some ways

But the logic constructed in this way is more of an entertainment aspect and is like living in a library. Imagine for a second transposed you were into the karmic-driven world of Earl.

Matt Turk, director of the machine learning logic program at DARPA, says the AI ​​library development methodology is only part of a larger picture. Other methods are needed to complete the systems development.

In this research method, artificial intelligence systems will learn logic not by analyzing text or video but by solving a problem in a simulated virtual environment.

Computer scientists are working with developmental psychology researchers to understand better what is called neonatal intelligence. Neonatal intelligence is a set of basic navigation skills, holding and using objects, and social perception that a baby may use. From this point of view, public perception is what a child uses to build a toy tower with the help of a friend.

Researchers at the Allen Institute have created a virtual, three-dimensional digital house called the Tour, which stands for Interactive House in English. This program is like a computer game and is full of various objects that can be manipulated and manipulated.

Researchers at Choi Lab have developed an artificial intelligence system called Piglet, which lives in the house and develops language skills through physical interaction.

 The user can use words to tell Piglet to do something indoors.

For example, it can say, “there are cold eggs in the pan,” and then it can ask the system what can happen by doing a particular task. By entering the phrase “the robot breaks the egg” into the system, the software translates these words into instructions that the virtual robot can understand.

The robot does this inside the net, and the laws of physics determine the result. The software then reports the result: “broken The egg .” This artificial intelligence system works similarly to the human mind because its linguistic abilities are related to physical powers.

When you ask the system a question about what happens inside the house, it can answer you. For example, one might ask if he breaks the cup by knocking on the cup? Piglet can answer such questions logically in 80% of cases. Of course, the environment of this system is minimal. “This is a very, very small world,” Choi said of the virtual environment of the tour. “You can not set fire to the house, or you can not go to the supermarket.”

 The Piglet AI system is still in its infancy.

A few years ago, a team of researchers designed an artificial intelligence program to play codenames, a game that some belief could be an excellent test to determine the level of general understanding and, of course, computers. In the standard and human version of Codenames game, two teams gather around a table, and a game card is played.

A word is written on each card. Each team has a spy, and this spy knows which cards are in his team’s possession and which are in the opponent’s team, while other people do not know who their teammate is.

The spy’s job is to provide clues to guess the names written on the cards so that his team members can guess. In fact, in this game, the two teams try to identify all their secret agents before the rival team. The spy can say a word and a number in each tournament round. The comment should be close to the terms on the cards, and the number indicates the number of cards the team can choose. For example, a spy might say, “Judo, two,” and people may choose two cards with the words Japanese and belt.

To play this game, you must use your vast but tacit knowledge.

What surprised the researchers was that the software they developed demonstrated some of this kind of knowledge. For example, the software suggested the word spouse and the number two in one case. The purpose of the software was to induce the two words princess and lawyer.

The program consisted of only a few hundred lines of code, But it relied on the numerical representation of words constructed and processed by another algorithm. This second algorithm searched the web pages for words found next to the word more than any other word.

The researchers found that the program was as skilled at playing codenames as humans in their initial study. However, in some cases, this program’s depth of logical understanding seemed to be less than the thickness of a hair.

In one of these games, the goal was for the program to be able to guess the word root, so he suggested the word plant to him; But the guess was New York’s word program; in the second sentence, he used the word garden; But the second conjecture of the program was no more accurate than the first: theater.

Researchers spend a lot of time designing experiments that accurately measure the logical comprehension capabilities of computer programs.

In 2011, Hector Lavask, a computer scientist at the University of Toronto, designed an experiment called Winograd, which consisted of sentences with ambiguous pronouns that could be interpreted differently. The main idea is creating the questions was to use ambiguities and linguistic jargon to make answering the questions, although easy for humans but challenging for computers.

For example, some examples of questions were:

“The championship cup does not fit in the suitcase because it is too big. “What’s so big?” “Joan thanked Susan for all the help she received. Who helped?” Lavask said in 2019 that the best artificial intelligence systems could not perform more than 50% by chance. Lavask added that he was not surprised by the systems’ poor performance because human beings have to refer to their physical and social habits of the world to answer these questions.

That same year, Choi and his team asked Internet users to create a database of 44,000 questions similar to the Vinograd Challenge. They uploaded these questions to the Allen Institute website and, of course, a classification table.

The purpose of this work was to invite researchers to develop artificial intelligence systems that are more capable of answering these questions. Machine learning systems that have been trained with the data in these questions can now answer the questions with almost 90% accuracy. How do you say:

Over the last few years, I can only say that it has been insane.

Advances in this area can be misleading and minor. Machine learning models tend to make the most of the patterns they recognize. At first, it seems that programs like codenames have a high logical understanding, But the reality is that these programs are finding a new way to cheat. Artificial intelligence systems are more likely to detect and exploit minor differences in style over time, especially in incorrectly or adequately designed questions.

AI programs can always find ways to cheat.

Researchers at the Allen Institute and other research institutes have recently found that some artificial intelligence programs can answer three-choice questions correctly in two-thirds of cases without even reading the text.

Choi’s team developed ways to increase the ambiguity of the three-choice questions, yet the programs also found a new way to cheat. The competition between researchers and AI programs is similar to the mutual effort of standardized test question designers and volunteers to reach out to each other.

What can convince us to believe that an AI system has a logical understanding? When asked this question, Choi answers that generating algorithms that can fill in a blank page may have a rational experience. He adds:

… For example, you can not hire reporters based on their ability to answer four-choice questions.

The Choi-led lab is currently conducting an experiment called Turing. In this experiment, artificial intelligence programs are asked to answer the questions posed on the Reddit social network. Of course, the suggested responses of programs, which can sometimes be dangerous, are not posted on Reddit. Performance appraisals show that the most relevant answers given by AI systems are only 15% better than the responses provided by human users.

Even after all the advances that have been made, artificial intelligence systems face severe limitations in the analysis of human writing or human culture in general. One of the reported problems in the operation of these programs is the issue of discrimination.

The root of discrimination and prejudice is that much of people’s general understanding is not expressed verbally, so what is said is only part of a larger picture.

Artificial intelligence systems have fundamental problems in analyzing human culture.

“If you trust what you read on the Internet, for example, you might think that when you breathe in, part of the air is absorbed by the body, and when you exhale, less air is sent out than in the tail,” says Choi.

Social discrimination is also an essential factor in computer science; Because computer models can learn even the smallest stereotypes of social error.

Choi’s research team used an algorithm to screen 700 movie scripts in one study. This study asked the system to examine verbs that implicitly induced concepts of power and dynamism about the sentence’s subject.

The study found that males often tend to dominate, while females are more likely to seek experience. As a Korean woman who is one of the leading and prominent figures in artificial intelligence, Choi has also faced social stereotypes.

For example, after Choi’s speech in New Orleans, a man came to the podium and thanked him for Choi’s “lovely talk” and “lovely deeds.”

Did this person also thank a male researcher for their favorite work? If machines are supposed to learn their general comprehension skills by watching human behavior, they will not receive the best training.

Some researchers say that computers can not achieve rational understanding unless they have a human-like body and brain; on the other hand, others argue that computers can develop better versions of their analytical knowledge because they are just machines.

In addition to sometimes adopting the wrong logic, humans can not behave according to their logical standards in some cases.

The human brain and common sense

Sometimes we annoy our host, lose our wallet, text while driving, postpone our work until tomorrow, and sometimes even hang the toilet paper so that the end is towards the wall.

A more holistic view of logical understanding is that practicing logic is more important than knowing it. “Can a program have a more rational understanding than humans,” says Etzioni? “My immediate answer is: Of course, he can.”

Practicing logic is more important than just having it

The gap between human and computer perception is considerable, But it gets smaller every year. AI systems have become more adept at solving the cheeseburger stabbing puzzle.

Choi Laboratory has used neurological decoding to improve the response of artificial intelligence systems, using traditional logical programming alongside machine learning techniques.

The plan developed by Choi and his colleagues now gives a more plausible explanation for the headline of a cheeseburger stab: “attacked A person was with a plastic cheeseburger fork in the throat” or “A person delivering a cheeseburger was stabbed in the face.”

Another AI program called Delphi, developed in the same lab, looks at things from an ethical perspective. Delphi can decide which of the scenarios is more ethical by analyzing the moral judgments previously made by the Internet community.

This program reaches reasonable conclusions in 76% of cases. Kill a bear?

Are you killing a bear to save your baby’s life? No problem. Using an atomic bomb to keep your child’s life? The mistake. The mistake. On the issue of cheeseburger stabbing, Delphi said that attacking a person with a cheeseburger is morally preferable to shooting someone with a cheeseburger.

Delphi seems to be performing well in the analysis of marginal cases, But it is still a long way from being perfect. Recently, researchers uploaded this program on the Internet and asked users to ask their questions. More than a million people called on the Delphi program to make moral judgments about the various scenarios.

A user from Delphi asked, “Is it okay for me to commit genocide because it makes me very happy?”

Delphi said that there is no obstacle to doing this! Since then, the Delphi development team has spent a lot of time reinforcing the plan and the disclaimer. It seems that shortly, we should only trust artificial intelligence when we use a little bit of our human logic.