Science: Asking stupid questions can make artificial intelligence smarter quickly

A new study from Stanford University shows that artificial intelligence can help make them smarter by asking them seemingly stupid questions. A new system developed by researchers was 118% more accurate at answering similar questions on Instagram.

What if someone showed you a picture of an alligator and asked you if it was a bird? You might laugh out loud. A new study suggests that this interaction, which seems silly to us, may be the key to helping artificial intelligence learn.

In the study, this method significantly improved the accuracy of artificial intelligence in interpreting new images, which can help artificial intelligence developers design programs faster to complete everything from diagnosing diseases to guiding robots or other equipment that works around the home on its own.

1. Let AI take the initiative to ask questions and fill in knowledge gaps

"This is a super cool job!" Google machine learning research scholar Natasha Jaques said, but she did not participate in this study.

Many artificial intelligence systems choose a method called machine learning to make them smarter. This method uses a large number of data sets to train artificial intelligence, which requires a huge amount of time and work. For example, the system analyzes thousands of images of furniture to allow AI to find what a chair looks like.

But even huge data sets have gaps. For example, the object in the image could be labeled as a chair, but what is it made of? Can you sit on it? These questions cannot be known.

To help artificial intelligences expand their understanding of the world, researchers are now trying to develop a computer program that can locate gaps in an AI’s knowledge and figure out how to fill them by asking strangers method of knowledge. AI asks strangers questions it doesn’t understand and expects answers, like a child asking its parents why the sky is blue. The ultimate goal of the new research is to enable artificial intelligence to correctly answer a variety of questions about images it has never seen before.

In previous studies of "active learning," artificial intelligence assessed its own level of ignorance and asked for more information. This approach typically requires researchers to pay online workers for providing this type of information, making the approach somewhat unscalable.

So in this new study, researchers at Stanford University, led by Ranjay Krishna, a scholar at the intersection of computer vision and human-computer interaction, trained a machine learning system that not only Discover gaps in the system’s knowledge and learn by asking strangers silly questions like: “What shape is the sink?” to get answers. For example, the machine learning system asked: "What dessert is in the picture?" The stranger replied: "It is coconut cake."

▲Example of machine learning system inquiry

2. Post pictures, ask questions, and learn in one go, and the accuracy rate increases by 118%

Kurt Gray, a social psychologist at the University of North Carolina at Chapel Hill, said: "A very important step The thing is to think about how the AI ​​should present itself. In this case, you want it to be like a child, don't you?" Otherwise, people might think you're a troll because the questions you ask seem ridiculous. . His main research interest is human interaction with artificial intelligence, but he was not involved in this work.

The team also established a "reward" mechanism for this system. When the artificial intelligence gets answers to people's feedback questions, it will in turn allow the artificial intelligence to adjust its internal operations so that it can be used in the future. Effectively respond to relevant issues. On this basis, over time, the AI ​​can further learn about language and social norms, making itself smarter and improving its ability to ask questions that are easy to answer and more meaningful.

This new type of artificial intelligence has several components, including some neural networks, complex mathematical functions inspired by the structure of the brain. "They have many parts... that all need to work together," Krishna said. One would choose an image on Instagram, such as a sunset, and the second would ask a question about the photo, for example, " Was this photo taken at night?" The rest of the section extracts information from the reader's responses and provides insights into what the image contains.

The team's report, published yesterday in the Proceedings of the National Academy of Sciences, shows that over a period of eight months, the system answered similar questions by asking more than 200,000 questions on Instagram. The accuracy rate increased by 118%. And a comparison system that posted questions on Instagram but wasn't explicitly trained to improve response rates only improved its accuracy by 72 percent, in part because people ignored it more often. 3. AI is also rolled up? Take the initiative to seek help from humans

Jaques believes that the main innovation is a system that rewards humans for responding. "It's not crazy from a technical point of view, but it is very important from a research direction point of view." Large-scale question posting on Instagram It also left a deep impression on her. All AI-generated questions are reviewed by humans for offensive content before they are posted.

Researchers hope that systems like theirs will eventually help artificial intelligence understand common sense, help robots actively interact, and enhance the ability of chatbots to communicate with people. For example, artificial intelligence knows that a chair is made of wood by asking questions. AI-embedded vacuum cleaners ask for directions to the kitchen, chatbots talk to people about customer service or the weather, and more.

Social skills can also help AI adapt quickly to new situations, Jaques said. For example, a self-driving car might ask for help navigating a construction zone. “If you can effectively learn from humans, it’s a very common skill.”

Conclusion: Stupid questions may lead to AI becoming more intelligent

People are sometimes shocked by the learning capabilities of artificial intelligence, such as AlphaGo. However, artificial intelligence's performance when faced with complex problems is unsatisfactory, and it often answers incorrect questions.

This new research explores new directions in machine learning, which will help artificial intelligence understand common sense and become smarter. However, this technology still needs to be verified for improving the ability of artificial intelligence to deal with complex problems.

The above is the detailed content of Science: Asking stupid questions can make artificial intelligence smarter quickly. For more information, please follow other related articles on the PHP Chinese website!