Main menu


When AI asks silly questions, it quickly gets smarter.chemistry

If someone showed you a picture of a crocodile and asked if it was a bird, you might laugh. And if you’re patient and kind, help them identify the animal. In a new study, this strategy dramatically improved the AI’s accuracy in interpreting new images. This approach could help AI researchers more quickly design programs that do everything from diagnosing diseases to guiding robots and other devices around your home.

Natasha Jaques, a computer scientist at Google who studies machine learning but was not involved in the research, said:

Many AI systems get smarter by relying on a brute force method called machine learning. For example, after analyzing thousands of furniture photos, find patterns in the data to understand what a chair looks like. But even huge datasets have gaps. Sure, the object in the image is labeled a chair, but what is it made of? And can you sit on it?

To help AI better understand the world, researchers are now trying to develop ways for computer programs to find knowledge gaps and figure out how to ask strangers to fill them. is blue. The ultimate goal of the new research was an AI that could correctly answer a wide variety of questions about never-before-seen images.

Previous research on “active learning,” where AI assesses its own ignorance and asks for more information, often finds that researchers have to pay online workers to provide such information. was. That approach doesn’t scale.

So, in a new study, researchers at Stanford University (now at the University of Washington, Seattle), led by Ranjay Krishna, train a machine learning system to not only find gaps in knowledge, but also about images perceived by strangers ( (often silly) questions. Answer patiently. (Q: “What is the shape of the sink?” A: “It’s square.”)

Kurt Gray, a social psychologist at the University of North Carolina at Chapel Hill who studies human-AI interactions but was not involved in that work, said it’s important to think about how AI exists. says there is. “You want to be like a child in this case, right?” he says. Otherwise, people might think you’re a troll by asking seemingly silly questions.

The team “rewarded” the AI ​​for writing easy-to-understand questions. When users actually responded to their queries, the system received feedback telling them to adjust their inner workings to behave similarly in the future. Over time, AIs have implicitly learned the lessons of language and social norms, honing their ability to ask sensible, easy-to-answer questions.

Q: What kind of dessert is the dessert in the picture? A: Hello, it’s coconut cake. very delicious 🙂 R. Krishna others, PNASDOI: 2115730119 (2022)

The new AI has several components, some of which are neural networks, complex mathematical functions inspired by the architecture of the brain. “There are so many moving pieces… they all have to be played together,” says Krishna. One component selected her Instagram image (such as a sunset) and the second component asked a question about that image. For example, “Is this photo taken at night?” Additional components extracted facts from the reader’s responses and learned about the image from it.

After asking more than 200,000 questions on Instagram over an eight-month period, the system’s accuracy in answering similar questions to those posed by the system improved by 118%, the team said today. Proceedings of the National Academy of SciencesA comparison system that posted a question on Instagram but wasn’t explicitly trained to maximize response rate only improved accuracy by 72% because people ignored it more often. .

A key innovation, Jax said, was to reward systems that made humans respond. “It’s not that crazy from a technical point of view, but it’s very important from a research direction point of view.” She’s also impressed by her large-scale real-world expansion on Instagram is receiving (Humans checked all AI-generated questions for offensive content before posting them.)

Researchers hope systems like theirs will eventually help us make sense of AI (For example, we know that chairs are made of wood)interactive robotics (AI-powered vacuum cleaner asking for directions to the kitchen)and chatbots (conversing with people about customer service and weather).

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