The idea that computers, or even machines at all, can understand what you’re saying is one of the greatest ideas in technology.
In fact, one of Google’s most successful ads has used the premise to sell the idea that machines will eventually be able to “understand” anything you say.
Now a team of researchers at the University of Southern California has found that this “understanding” might be less accurate than it appears.
Their work, published in the journal Computers in Human Behavior, suggests that machines may be less able to understand the meanings of what people are saying, or to infer from the way we’re talking.
That’s not good news for people, as it means that machine-learning algorithms that can learn about the meanings we use to say things may end up being less useful than we thought.
The team’s work, however, was not limited to humans.
Their findings also extend to other groups of animals, and they suggest that humans might be particularly vulnerable to this type of “intrusive learning”.
And they might also have a bigger impact on the future of human-computer interaction than people realise.
“This work is a real game changer in the field of machine learning,” says Nicholas Fagan, an assistant professor at USC and lead author of the paper.
“It’s a great example of how a lot of progress in machine learning can be made by using techniques that were previously very difficult, but now have been very effective.”
The researchers used machine learning techniques to measure how quickly humans and animals can process words.
They then used the data to predict which words are the most likely to trigger human-machine interaction.
This revealed that animals are more likely to use “unrelated” words that trigger interactions, such as “sneezes” and “breathing”, which are more common in human speech.
In humans, the opposite is true: words that evoke a response, such in the phrase “I smell a rat”, are more often used.
And these words are often linked to animals, so the animals are also more likely than humans to be able and willing to use the word “sniff”.
However, in contrast to the human results, the animals were not able to learn to differentiate between “he,” which is common in speech, and “she,” which might be used in a conversation, and which is often used in animals as a sign of trust.
“What we found is that humans and other animals are far more similar than they are different,” says Fagan.
“The difference is in how the animals learn to use these terms.
Animals learn to be very good at recognizing unfamiliar words and phrases and then they learn to infer the meaning.
They learn to make a prediction and then use that to infer whether or not they want to communicate with the other animal.”
In humans and primates, humans and chimpanzees often have different patterns of neural activation when they hear unfamiliar words.
“Humans have a relatively large number of neurons in the ventral tegmental area, which is where the ‘I’ and ‘S’ neurons are located,” explains Fagan (he is also an assistant director of USC’s Center for Neural Computation).
“Our work shows that this area of the brain is particularly active when humans and monkeys use the same word to refer to each other.”
The same is true for other animal groups, such an elephant who uses the phrase to communicate trust, for example.
Fagan and his colleagues also used a more sophisticated machine learning algorithm called a “classifier” to determine the “intraspecific similarity” between the two human-animal groups, which was done with a combination of natural language recognition, video analysis, and fMRI data.
“In this work, we looked at how human-monkey pairs of monkeys and humans interact in the same environment,” says Dr. Mark Weiler, an associate professor at the USC College of Letters and Science and lead researcher of the study.
But humans were also less likely to say the word. “
When we use the words ‘I smell rats’ and then the word ‘sniff’, humans are much more likely [to] use the latter word.”
But humans were also less likely to say the word.
This suggests that human-specific similarities are not just a result of human language learning, but are more complex than that.
The researchers also looked at whether other animals could infer the context of human speech, but did not find evidence that humans were more effective at that task.
“Our results indicate that humans do not use the more generic ‘snipe’ and that monkeys use ‘disco’.” Humans also seem to be less likely than other animal species to use words that cause other animals to respond in a negative way.
“Given that animals can learn to respond to the same words, they also have learned to use those