Original article was published by Haucem Sadki on Artificial Intelligence on Medium
The second main branch of AI concerns human-machine interaction and in particular natural language processing.
For a machine, being able to answer a question asked by a human requires three steps: hearing what the user has said, understanding the meaning of their request, and responding to their expectations.
It is by using “machine learning” that the internet giants (Google, Facebook, Microsoft, IBM, Apple, Amazon…) have recently made enormous progress on the first step. By training their algorithms with the demands we all make on their systems, they have become as efficient as humans.
And then … Will the system be able to make sense of what the user is saying? Here things take a turn for the worse: current systems are unable to understand what they are being told.
If I ask the machine “Bring me a baguette”, it will be unable to “guess” an answer satisfying my request, because unable to perform an action that it has not been taught, such as, here, placing an order with my baker to have my bread delivered to me.
This limitation exists because the latter request cannot simply be associated with information. It must induce the execution of a series of actions (the third step).
And when it’s possible, like sending an SMS for example, it is because a human has associated the request with a function that the phone can perform. Such functions are not, to this day, invented by machines, but programmed by hand, by a human being like both of us.
We‘re the ones writing the script
So let’s not fool ourselves: in the current state of technology, the vast majority of actions performed by voice assistants are therefore not invented by machines but programmed specifically by computer scientists.
To claim that machines are able, through AI, to learn to do things on their own, has been a hoax to this day. As is claiming that machines can anticipate a situation that has never been encountered.
Let us agree that the law of large numbers and the digitization of all human activities allow classification to be made with great efficiency. We agree that this will greatly disrupt our daily life.
Let’s also agree that the gap between classifying the information and understanding it is (still) sizable. And let’s agree that how we deal with closing that gap will depend on the society we are heading to.
Can we combine technology and intelligence?
Artificial intelligence has capabilities that are similar to human intelligence. Honestly speaking, it‘s not able to think, but it can create, reproduce, and solve problems.
It can be trained, adapt, learn continuously, just like a human being tends to discover new things and practices them. With this understanding, it is possible to say that AI is indeed “smart”.
However, these comments need to be tempered: although in a certain sense, AI is “intelligent”, its learning process is entirely different from ours.
For a man or a woman, to learn to recognize a cat, he or she has to see the animal only a few times. For a program, it has to analyze thousands of images to classify it.
“We have built systems capable of recognizing cats with a 95% success rate by providing them with 100,000 images of cats. Whereas a child only needs two cat images to recognize one all his life, with a 100% success rate. »