Some of the greatest discoveries come not just from observation, but from thought. Einstein developed the theory of relativity through thought experiments, and Galileo gained insights about gravity through mental simulations.
A review published on September 18 in the journal Trends in Cognitive Science suggests that this thought process is not unique to humans. Artificial intelligence can also correct itself and draw new conclusions through “thinking and learning.”
“There have been some recent demonstrations of artificial intelligence, especially in large language models, that look like learning by thinking,” said author Tania Lombrozo, a psychology professor and co-director of the Natural and Artificial Thinking Program at Princeton University. “Sometimes ChatGPT corrects itself without explicit notification. This is similar to what happens when people learn by thinking.”
Lombrozo listed four examples of learning through thought in humans and artificial intelligence: learners can obtain new information through interpretation, simulation, analogy, and reasoning without external input. For humans, explaining to children how microwave ovens work may reveal gaps in our understanding.
Rearranging furniture in the living room often requires creating a mental image to simulate a different layout before making any physical changes. Downloading pirated software seemed morally acceptable at first, until someone compared it to stealing physical goods. If you know that a friend’s birthday is leap day and tomorrow is leap day, you can infer that your friend’s birthday is tomorrow.
Artificial intelligence exhibits a similar learning process. When asked to elaborate on a complex topic, the artificial intelligence may correct or refine its initial response based on the explanations it provides.
The gaming industry uses simulation engines to approximate real-world results, and models can use the output of simulations as input for learning. Asking a language model to make analogies allows it to answer questions more accurately than simple questions. Prompt artificial intelligence to conduct step-by-step reasoning can enable it to get answers that cannot be obtained by direct queries.
“This raises the question: why both natural and artificial thinking have these characteristics? What is the use of learning through thinking? Why is it valuable?” Lombrozzo said. “I think learning through thinking is a kind of ‘learning on demand.'”
When you learn new things, you don’t know how this information will serve you in the future. Lombrozzo said people can save this knowledge for later use until the context makes it relevant and worth the cognitive effort of thinking and learning.
Lombrozo acknowledges challenges in defining the boundaries between reasoning, learning and other higher cognitive functions, an area of debate in cognitive science. The review also raised additional questions, some of which Lombrozo plans to explore further, such as whether artificial intelligence systems are actually “thinking” or simply imitating the output of such processes.
“Artificial intelligence is already very complex in some ways, but limited in others, so we have the opportunity to study the similarities and differences between humans and artificial intelligence,” Lombrozzo said.
“We can learn important knowledge about human cognition through artificial intelligence and improve artificial intelligence by comparing it to natural thinking. This is a critical moment and we are in this new position to ask these interesting, comparative questions.”
There are links in the video below this video. If you are interested, you can open it and have a look.
Thank you for watching this video. If you like it, please subscribe and like it. thank
Original text:https://techxplore.com/news/2024-09-humans-artificial-minds.html
More information: Learning through natural and artificial thinking, Trends in Cognitive Science (2024). DOI：10.1016/j.tics.2024.07.007
Journal Information: Trends in Cognitive Science
Provided by Cell Press
Oil tubing: