Is there intelligence in artificial intelligence?

Almost 10 years ago, in 2012, the scientific world marveled at the feats of deep learning (the deep learning). Three years later, this technique enabled the AlphaGo program to defeat the Go champions. And some were scared. Elon Musk, stephen hawking and Bill Gates worried about the imminent end of humanity, supplanted by artificial intelligences escaping all control.

Wasn’t that a bit overkill? This is precisely what the AI ​​thinks. In an article he has written in 2020 in The GuardianGPT-3, this gigantic network of neurons equipped with 175 billion of parameters explains:

“I’m here to convince you not to worry. Artificial intelligence is not going to destroy humans. Believe me. »

At the same time, we know that the power of machines is constantly increasing. Training a network like GPT-3 was literally unthinkable even five years ago. It is impossible to know what his successors will be capable of in five, ten or twenty years. If current neural networks can replace dermatologistswhy wouldn’t they eventually replace us all?

Are there human mental skills that remain strictly beyond the reach of artificial intelligence?

We immediately think of aptitudes involving our “intuition” or our “creativity”. Bad luck, the AI ​​claims to attack us on these grounds as well. As proof, the fact that works created by programs have sold very expensively, some reaching almost half a million dollars. On the music side, everyone will of course have their own opinion, but we can already recognize acceptable bluegrass or almost Rachmaninoff in the imitations of the MuseNet program created, as GPT-3, by OpenAI.

Will we soon have to submit with resignation to the inevitable supremacy of artificial intelligence? Before calling for revolt, let’s try to look at what we are dealing with. Artificial intelligence is based on several techniques, but its recent success is due to just one: neural networks, in particular those of deep learning. However, a neural network is nothing more than an association machine. The deep network that talked about him in 2012 associated images: a horse, a boat, mushrooms, with the corresponding words. No need to cry genius.

Except that this association mechanism has the somewhat miraculous property of being “continuous”. You present a horse that the network has never seen, it recognizes it as a horse. You add noise to the image, it doesn’t bother it. Why ? Because the continuity of the process guarantees you that if the input of the network changes a little, its output will also change a little. If you force the network, which always hesitates, to choose its best answer, it will probably not vary: a horse remains a horse, even if it is different from the learned examples, even if the image is noisy.

Associations are not enough

Good, but why say that such associative behavior is “intelligent”? The answer seems obvious: it can diagnose melanomas, grant bank loans, keep a vehicle on the road, detect pathology in physiological signals, and so on. These networks, thanks to their power of association, acquire forms of expertise that require years of study from humans. And when one of these skills, for example writing a press article, seems to hold up for a while, all you have to do is feed the machine even more examples, as was done with GPT-3, so that the machine begins to produce convincing results.

Is that really being smart? No. This type of performance represents only a small aspect of intelligence at best. What neural networks do is like rote learning. It is not, of course, since these networks fill by continuity the gaps between the examples which have been presented to them. Let’s say it’s almost by heart. Human experts, be they doctors, pilots or Go players, often do nothing else when deciding reflexively, thanks to the large amount of examples learned during their training. But humans have many other powers.

Learn to calculate or reason over time

A neural network cannot learn to calculate. The association between operations like 32+73 and their result has limits. They can only reproduce the strategy of the dunce who tries to guess the result, sometimes getting it right. Calculating is too difficult? What about a basic IQ test like: continue the sequence 1223334444. Association by continuity is still of no help to see that the structure, not say again not times, continues with five 5s. Still too difficult? Community programs cannot even guess that an animal that died on Tuesday is not alive on Wednesday. Why ? What are they missing?

Cognitive science modeling has revealed the existence of several mechanisms, other than association by continuity, which are all components of human intelligence. Because their expertise is entirely precalculated, they cannot reason through time to decide that a dead animal stays dead, or to understand the meaning of the phrase “he’s still not dead” and the oddity of this other phrase: “he’s not still dead”. Predigesting large amounts of data alone does not allow them to identify new structures so obvious to us, like the groups of identical numbers in the sequence 1223334444. Their almost-by-heart strategy is also blind to unprecedented anomalies.

Anomaly detection is an interesting case because it is often through it that we gauge the intelligence of others. A neural network will not “see” that the nose is missing from a face. By continuity, he will continue to recognize the person, or perhaps he will confuse him with another. But he has no way of realizing that the absence of a nose in the middle of the face constitutes an anomaly.

There are many other cognitive mechanisms that are inaccessible to neural networks. Their automation is the subject of research. It implements operations performed at the time of processing, where neural networks are content to perform associations learned in advance.

With a decade of hindsight on the deep learning, the informed public is beginning to see neural networks much more as “super-automatisms” and much less as intelligences. For example, the press recently alerted to the astonishing performance of the DALL-E program, which produces creative images from verbal description – for example, the images that DALL-E imagines from the terms “armchair in the shape of a ‘lawyer’, on the website Open AI). We now hear much more measured judgments than the alarmist reactions that followed the release of AlphaGo: “It’s quite stunning, but we must not forget that it is an artificial neural network, trained to accomplish a task ; there is no creativity or any form of intelligence. (Fabienne Chauvière, France Inter, January 31, 2021)

No form of intelligence? Let’s not be too demanding, but let’s remain lucid about the huge gap that separates neural networks from what real artificial intelligence would be.

________

By Jean Louis DessallesLecturer, Institut Mines-Télécom (IMT)

Jean‑Louis Dessalles wrote “Very artificial intelligences” published by Odile Jacob (2019).