The spirit of geometry and of finesse

Artificial intelligence: current state, development, challenges and risks

by Claude Roche, former Vice-President of the Academy of Air and Space- Paris, Board member of EuroDéfense-France

Published in THE EUROPEAN – SECURITY AND DEFENCE UNION - Volume 50 - 1/2024

The official start of Artificial Intelligence (AI) goes back to the Dartmouth Conference (US) in 1956, with Marvin Minsky, John McCarthy and Claude Shannon, where the expression was first coined, at a time when computers were becoming more widespread.

At the outset, researchers' concerns were focused on machine learning, with game theory being applied, in particular, to chess, theorem proving, operational research, image processing, recognition of pictorial patterns and speech. Simultaneously, computing power was increasing considerably, and in 1997 the Deep Blue computer beat the world chess champion, G. Kasparov.

Since then, pattern recognition and symbolic AI have continued to merge and AI has taken off. Intellectuals, businesses and the media swear only by AI, predicting that in just a few decades, the machine will surpass man!

What is a pattern or a concept?

Patterns and concepts exist when there are facts that are not purely random and that relationships therefore exist with other patterns or concepts. The patterns and concepts then constitute an immense Russian nesting doll structure, the lowest level of which is the output of optical, acoustic or other sensors and show- ing total continuity for patterns, concepts, ideas, theories, etc. The whole provides a representation of the world outside man, a living being or a machine.

Pattern recognition and symbolic AI

• Pattern recognition studies the detection and description of patterns or concepts by machines to satisfy different societal or industrial needs.

• Symbolic AI studies how to solve problems based on these patterns or concepts, giving rise to particular programming difficulties.

Pattern recognition can be/is used in most symbolic systems as a means to choose between the “good choice” or the “bad choice” of patterns compared to the others. These systems or methods are nowadays called “hybrid AI”.

The functions of human intelligence: bijection between reality and memory

The aim of the two intelligences is to combine to obtain the best possible bijection between reality as humans apprehend it and the content of their memory, by allowing them to influence both. The medium for these functions is memory.

In the medium term, AI will only partially cover “life” intelligence (self-driving cars, household robots, security systems) and not “invention/creation” intelligence, except when this partially touches on intuition through deep learning.

The strict definition of AI

When the computer must make decisions alone and activates a life loop and when

1. the solution is defined entirely by a logical model: this is programming.

2. the problem cannot be solved by a rigorous solution or the computer cannot apply it fast enough, but it is desirable to find a solution acceptable in practice: this is artificial intelligence.

While programming produces model errors and bugs, AI, in addition to bugs, always has a probability of error due to the absence of a rigorous model: GPS, voice and facial recognition, etc.

The creation of AI, the logic of development

To create artificial intelligence, developers must, based on their intuition and knowledge of the problem, search, randomly or by deduction, for functions that can be assembled in multiple ways and evaluate these assemblies on a set of “learning data” representative of the phenomenon, and above all, repeat the process many times over. Finally, they must validate the result on other representative data and, if the performances are satisfactory, validate the resulting system in operation.

Today, there are numerous development experiences in extremely diverse fields, numerous ideas for methods and tools, such as those linked to big data and clouds, probability and correlation calculations with the use of statistics, graph paths with minimax, branch and bound, game theory, image and sound processing, deep learning, and all models with adequate processing power such as 2D, 3D, 4D, syntax and semantics for speaking and reading, etc.

The most important tool remains the intelligence of the individual researcher-developer who assembles all these bricks, on the principle of “trial and error” or “test and learn”. Indeed, current AI developments are still mainly human-driven, with only a marginal and complementary input from machine learning. And they only involve the “life loop” functions, and not yet by any means the symbolic intelligence of the “invention/ creation” loops.

Let us note in passing that this “trial and error” principle is also the basis of evolution: of the cosmos, the Earth, humanity and deep learning! And even, according to Stephen Hawking, of the Big Bang itself!

AI development issues

1. Basic risks due to the insufficient quality of developments

• The goals of the constructed system may not have been sufficiently well defined: either they are not complete, or the system might find itself in a situation where two goals are contradictory, like the “psychological” problem of the HAL machine from “2001: A Space Odyssey”.

• Training and testing data might not be representative of important exceptional situations.

• All AI systems have a non-zero error rate. Human beings have exactly the same faults!

2. Societal risks of developed systems

• Civil aviation requires a level of reliability that AI cannot possibly attain.

• Before the car accident, should the AI decide to kill other road users or the family in the car?

• What about autonomous military robots, when, according to the code of war, we have no right to seek to kill if we are not sure of our target?

• We demand perfection from a robot that we cannot prosecute in court.

All sensitive systems will take a lot of time and adaptation before they can be used, unlike other systems such as face recognition, language understanding, etc.

3. Man-machine interfaces

• Operational relationships between man and machine will have to adapt, in the same way that human pilots adapt to autopilots.

• Man-machine interfaces are not yet perfectly resolved. By introducing a third player, how will the system cope with three times as many interfaces?

• A huge number of simulations and time-consuming experiments will be necessary, with intense and multiple reflections.

Research to be carried out before unleashing strong AI

Strong AI will be a much bigger step than the one which has led to our current AI.

• Its memory will be many orders of magnitude greater than that of current AI,

• It will learn to learn and test its efficiency in view of better certification.

• It will invent and create, with the “invention/creation” loop that current AI does not have.

Getting there will take many decades of research, going much further than deep learning.

It will be necessary in particular to teach the machine to discover a new pattern or concept, by detecting in the observed universe that something is out of the ordinary and triggering a search to define what it is by symbolic representation. We will also use the recursion of operators: the methods for discovering patterns and concepts will themselves be based on concepts, which can be discovered using the same method.

It is then, and only then, that we will be able to imagine AI systems approaching human intelligence, whose functioning we will be able to formally represent and certify, whereas today even deep learning creates systems whose rigorous decision-making we cannot understand. We will thus have taught computers to learn intuition and rationality, both referred to as “spirits” by Blaise Pascal: the spirit of geometry and the spirit of finesse.

------------