(ORDO NEWS) — American scientists have presented a new mathematical model that can improve AI performance by combining human and algorithmic predictions and confidence estimates.
Separately, human and artificial intelligence periodically fail in different tests, but their combination can lead to the creation of the first strong AI.
From question-answering chatbots to algorithms driving autonomous vehicles and making medical diagnoses, artificial intelligence underpins many aspects of everyday life. But so far, many of the uses we’ve seen in science fiction remain elusive because it’s not possible to endow AI with some of the power of human thought.
How to make AI more human
Human and machine algorithms have complementary strengths and weaknesses. Each of them uses different sources of information and strategies to make forecasts and make decisions.
Using theoretical calculations and experiments, the researchers showed that humans can improve AI predictions even when human accuracy is somewhat lower than AI accuracy and vice versa. And this accuracy is higher than combining the predictions of two people or two artificial intelligence algorithms.
To test this concept, the researchers conducted an image classification experiment in which humans and computer algorithms worked separately from each other.
They had to correctly identify distorted images of animals and household items – chairs, bottles, bicycles, trucks. Humans and AI rated their confidence in the accuracy of identifying each image as low, medium, or high.
The results showed large differences in recognition accuracy between humans and artificial intelligence algorithms in this test: in some cases, people were highly likely to claim that the picture was a chair, while the algorithm was confused, in other cases the opposite was noted.
When human and computer predictions and confidence estimates were combined, the hybrid model performed better than either human or machine predictions alone. Roughly speaking, scientists have managed to achieve synergy between man and machine. These advances could help develop better human-like algorithms.
—
Online:
Contact us: [email protected]
Our Standards, Terms of Use: Standard Terms And Conditions.