AI: right or wrong?

ANN (artificial neural networks) + deep learning = artificial intelligence that could grow through training and the consumption of massive quantities of data. But can a lot of data really be equated with intelligence?

ANN can’t tell the difference between a photo of a cat and a photo of a dog, unless it has been trained to do so by analyzing an enormous cache of photos of the two animals. That’s worth repeating: ANN needs a great deal of data and training to recognize anything, and even after all that, it won’t be able to explain why it made a particular choice. This is the “black box” effect of the all-powerful ANN + deep learning combination: although the data, like the conclusion, is known, the AI decision-making process is entirely impenetrable.

And it’s often accompanied by a mantra that is being challenged more and more: “It’s the correct answer on average”. So, what would you do if your AI, after it has been trained, still can’t distinguish a puppy from a cat? The temptation would be to force-feed it new data. But would that make it any smarter? Thomas Solignac, co-founder of Golem.ai, draws a neat parallel: “It’s as though, if you wanted to teach a child how to read, you just left them alone with hundreds of books.

And, when they failed, you decided to add a few hundred more!” Plainly speaking, we shouldn’t confuse huge volumes of data with intelligence: the most efficient algorithms know how to select the right information to make a decision. When will AI be more like a human brain, rather than a super statistical analyst? There are those who predict it will take 20 to 30 years to reach that point, while others say we’ll only have to wait until 2020.

Find out more:

Pourquoi le machine learning n’aura jamais tort” February 2019 and “Il n’y a pas une once d’intelligence dans l’IA” (Usbek & Rica, June 2018)