The revival of AI is undeniable and there is no day left without a spectacular announcement on a particular success.
These technological advances are so blatant that AI has become a strategic issue: in France the latest report of the Academy of Technology , and the Villlani report (and many others in the US/Anglo-Saxon sphere) advocate for companies and the state not only to take a close interest in them but to implement them.
If the enthusiasm and generalization to all areas of human activity followed immediately, the most disturbing speculations about the future of employment and hazardous on the future of humanity were also developed.
But … looking at it more closely:
- Chatbots are capable of a simple conversation in a very limited context ( booking a restaurant ),
- but remain unable to understand the complexity of the world (normal because they do not live there!), and especially have a real problem of learning (see the racist Tay Robot of Facebook ).
- Translation systems are statistically very interesting and practical,
- but fail at the smallest grammatical complication or variant in languages (example: ” he put on the right his towels and on the left her towels ” (english sentence) is translated to french by Google in: “« il a mis à droite ses serviettes et à gauche ses serviettes »” (“he put on the right his towels and on the left his towels”) which does not mean anything …).
- The AlphaZero program learns to be world champion of strategic board games (like Go and Chess games),
- but stumbles on an elementary game that he can not solve … and takes no pleasure in playing!
- AlphaZero learned in a very short time (a few days) thanks to a colossal computing power.
- IA software provides, in advanced cases, a useful assistance to the diagnosis of cancer by image processing ,
- without understanding the basics of cell biology (so back to basic expert system of the 90s !!!) and even less medicine.
- The programs of artificial vision recognize objects already seen in images,
- but do not know how to generalize as the human mind does, and moreover they are easily fooled .
- Autonomous car driving programs arrive on the roads,
- but still have great progress to make: for example, in France, the rate of fatal accidents is 1 per 100 million km traveled. The demonstration that smart vehicles do better requires traveling 1 billion kilometers! Should we proceed by trial-and-error and how long for how many deaths in the US with ‘stupid’ versions?
- Classification algorithms are very practical in marketing, for which the impact of an error is marginal,
- but have big bias problems in other applications.
- Japan’s Todai robot passes university entrance exams better than 80% of students (video) ,
- whereas he proves incapable of answering simple questions of common sense (cf 8:34).
In the end, AI’s successes are confined to very limited tasks. These programs are only an illusion of intelligence: they are all very specialized, very constrained to the problem they solve, unable to generalize and all very sensitive to the elements that allowed their learning.
A generic problem?
A long article by Gary Marcus , a professor at New York University offers a similar analysis, while the most advanced researchers in the field like Yann le Cun also warn about the limits. Going even further, the famous AI scientist Roger Shank illustrates in a recent paper the marketing strategy of IBM on the AI that he even qualifies as ‘ fraudulent’. A recent article from the Google Labs describes the limitations currently seen in AI medicine, which are not far removed from those of the 90s!
All these cases demonstrate that current fashioned methods for deep-learning, which are only very analytical, are far from sufficient to represent the complexity of human intelligence and various forms of reasoning, including almost immediate abilities to learn and generalize. Even if the frontier has been pushed, the paradigm of generic AI is therefore still as present as during the last winter of AI!
Human intelligence is very broad and has allowed it to understand millions of topics (Wikipedia contains nearly 20 million articles in English), and to continuously adapt to new areas, while the successes of the IA covers only a few very specialized areas, with glaring limits and huge human and technical efforts.
It is therefore to be very cautious and realistic in the implementation of ‘magic’ IA tools: the hypothesis “an algorithm for all” (= neural network and variants) does not work properly even with important adaptations ( LSTM ). The sensitivity of these softwares to the calculation in double or simple precision shows us an incompatibility with the human reasoning which does not have to do of the 20 ° decimal of a reasoning.
Finally, the manufacture of fake thanks to AI is a booming industry that needs to be urgently addressed as it poses an immediate danger to humanity and freedom : the new AI brings in more problems than solutions !
The considerable progress that we see is due in large part to a colossal increase in computing resources since the 2000s, as well as to algorithmic and conceptual advances in neural networks, but fundamental problems remain, such as the understanding of natural language, understanding of the world, understanding of human knowledge, common sense, etc.
A new danger for AI, its implementation and research in the field
The journalistic frenzy and the financial and advertising stakes of the main actors should not lead us to a new “AI winter”, generated by unrealistic expectations, punctual investment returns or implementations with serious consequences.
On the one hand, the machines and the software have evolved considerably these last years, and one should not hesitate to acquire competences and to realize prototypes in this domain.
However the teams put in place must have much broader skills than statistics and the implementation of big data, neural networks and TensorFlow. Critical and detailed implementation must be studied carefully before any deployment.
The question “what can I do with these algorithms” should be doubled by a series of questions:
- what can I not do with these algorithms?
- what are the limits ?
- what are the consequences of an error (minor, repairable, …)?
Oddly enough, most of these basic questions are forgotten in the enthusiastic and populist statements. Progress, however, comes from the impossible today.
There is clearly still a lot of room for research , and we must get out of this mode “any application to its network of neurons that must be trained” to expand the techniques used. AI without neural networks has found a huge range of applications in decision support systems, CAD tools, Mathlab, credit card control and other software. The scientific spirit and the reasoning allowed original developments by widening the spectrum of knowledge beyond the habits of the old. Aircraft have been built while birds have neither propeller nor reactor. As long as modern AI is confined to a statistical (even astute and mathematical) review of facts, it can evolve only within its own narrow limits.
© Laurent Gouzènes, May 2018
KM2 is a generic tool for managing and manipulating knowledge for Augmented Intelligence.