Researchers have found that when neural networks with different training labels are compared, they perform better.
Binary language is compact and accurate fortransmission of information. In contrast, spoken human language is more tonal and analogous. Because numbers are an efficient way to digitize data, programmers rarely use other types of inputs when designing a neural network.
One of the most common exercises fortesting a new machine learning method: teaching AI to recognize objects or animals in a photograph. The authors of the new work conducted an experiment: they created two new neural networks that were supposed to recognize ten different types of objects in a collection of 50 thousand photographs.
The first AI system was trained in the traditional way: it was loaded with a data table of thousands of rows, each corresponding to one training photo.
And the authors loaded the table into the second systemdata, the rows of which contained a photograph of an animal or object, and in the second column there was an audio file in which a person pronounces the name of the object or animal.
As a result, the first neural network produced digitalthe meaning of the object that was shown to her, and the second tried to “tell” what she saw. Both algorithms coped with the task equally efficiently and answered correctly in 92% of cases, the authors note.
However, the results of the experiment changed whenscientists reduced the sample from 50 thousand to 2.5 thousand. Then the correctness of the answers of the first AI dropped to 35%, and for the second, which was trained by voice, it dropped to only 70%.
Read more:
Researchers plunged for the first time to the deepest sunken ship
The first accurate map of the world was created. What's wrong with everyone else?
A wireless system has appeared that helps paralyzed