No one can now ignore the near omnipresence of artificial intelligence. It is everywhere, from our daily sports or diet applications to the decision support systems of the world's largest companies, via visual or semantic recognition platforms.
AI, AIs, today represents a multi-billion euro industry, growing extremely dynamically. It radically changes the way we live and work. It impacts the organization of companies, their operations, the skills they seek. And this transformative movement is only at the beginning.
And yet, we cannot overlook the fact that many point out that AI today does not really exist, in particular Luc JULIA, the co-creator of SIRI. And indeed, the vast majority of AI systems at work to date:
- are more automation systems, decision trees, linked to heavy work natively biased and incomplete,
- or depend on a statistical calculation which limits its relevant use to phenomena that are in the strong majority - those that ensure a probability of satisfactory reliability - and is unable to account for what is in fact a large majority: a myriad of micro phenomena,
- impose an enormous work of definition and labeling of categories, carried out by… human armies. Remember the opinion critics a few months ago, complaining that Alexa’s teams were "listening" to us. Obviously. She needs it to progress.
Yes, AI is already powerful, but its true potential has yet to be unleashed. Far from the fantasies of “good conscience” or fashionable questions about ethics.
In a very interesting article published four years ago, Scott Jones (co-founder of Six Kin Dev. and Lecturer at the Singapore Institute of Technology) distinguished 4 waves of AI:
The 1st wave, which he places from the 70s to the 90s and which he calls GOFAI (Good Old Fashion AI) developed AI based on rules set by humans. He is talking here about decision trees. Able to apply the designer's will but unable to learn or generalize. They still constitute the vast majority of AI systems in use today, and in particular decision support systems. They are still mainly used in terms of personalization or customer service.
The 2nd, the current one, from the 90s to the present day, is, according to him, capable of learning but limited in reasoning and in the ability to generalize. This is the wave of machine and deep learning. Based on statistical learning and using “deep” neural networks, they are particularly used in advanced text, speech, language and vision processing. And, wrongly, when it comes to personalization.
His opinion is clear: “Current commercial AI technology falls squarely into the category of 'narrow AI', i.e. highly specialized systems that are very good at specific, well-defined tasks... and nothing else. Even self-driving vehicles, impressive as they are, use a tight set of AI systems. If you took software from a self-driving car and put it in a golf cart, it would be useless without reprogramming. In contrast, any human who had learned to drive a car could ride a golf cart for the first time and have no problem navigating the fairways. This is of course because humans are very good at abstraction - we can easily generalize the solutions and apply them to similar but different problems. Contemporary AI systems cannot do this.”
They simply know how to deduce and not induce like the human brain.
He then describes a 3rd wave, which will implement systems that will be capable of perception, reasoning and, above all, generalization. That is to say the ability to learn from a single experience, like a human, and not from the emission of a probability by statistical processing of a large number of occurrences.
For him: “Third-wave AI systems will show dramatic improvements, especially in their ability to adapt to context. They will understand the context and meaning, and be able to adapt accordingly. Third wave AI will not only recognize a cat, but will be able to explain why it's a cat and how it came to that conclusion - a giant leap from the "black box" systems of today."
And he even imagines a 4th wave, “able to accomplish any intellectual task that a human being can accomplish”. He sees it as an Artificial General Intelligence eventually leading to Artificial Superintelligence and “Technological Singularity”. Maybe.
Netwave's Inductive AI, patented in 43 countries, belongs to this 3rd wave of AI.
Unlike the systems of machine and deep learning, its inductive “reasoning” system allows it to learn from a single experience and generalize from it. This is fundamental when it comes to real-time digital interaction.
The most important leap is undoubtedly this ability to generalize that machine learning lacks. On the web, whether it's commerce, content, advertising, you can't wait for a behavior to become dominant in your learning base to generalize it.
You will lose months of sales, reading, clicks.
A brief reminder of the notions of induction and deduction:
When you deduce, you apply pre-defined rules to pre-defined special cases. In the case of 2nd wave narrow AI systems, the pre-definition of rules and cases is done by more (or less) periodically updated statistical processing.
When you induce, you generalize what you experienced in 1 particular case. And you give this generalization a relative value: you agree to question it when you live a different experience. You could say that Inductive AI is agile AI. She works by iterating and constantly questioning herself.
To be complete, we would have to add a third cognitive mechanism: abduction. To put it simply, it's the same thing as induction but, instead of giving a relative value to the lived experience, you give it an absolute value: “all mechanics are thieves”... A thousand apologies to the mechanics or the children of mechanics who will read me but the example will speak.
The advantages of Inductive AI, which is able to generalize from an experience, in terms of personalization in the web environment, are spectacular:
- Rather than contenting yourself with a few dozen pre-defined segments - in the past - to “describe” your visitors, Netwave AI identifies 85,000 different situations every 1 million visits and manages to exploit 16,000 of them.
- You are no longer in the automation of pre-defined rules but in a real individualization,
- You identify new situations in real time, where you would have had to wait for a new processing to detect them.
- Real time allows you, not to presuppose the need of your visitor but to really identify it and to be able to take into account the evolution of your visitor between two visits and during the visit itself.
- You identify the needs of the “long tail of behaviors” those which, as distinct groups, individually concern few visitors, but whose sum can represent up to 80% of your purchases. That is to say the main part of your turnover…
- Rather than applying to the visitor the rule applicable to the segment to which you have attached it, resulting from a statistically pre-calculated probability, you generalize what you have experienced most recently in the same situation.
- The granularity of the situation analysis (16,000 situations every 1 million visits) allows you to dramatically increase the relevance of your personalization,
- Generalizing from the most recent identical situations allows you to reproduce new trends, emerging behaviors, immediately, without wasting weeks or months waiting for them to have acquired a majority position in your learning base.
- You are no longer dependent on huge sets of training data. Do you know that a typical machine learning algorithm dedicated to dog recognition, for example, has to be fed tens of thousands of dog images before reaching an acceptable level of accuracy, which a child of three years can reach after seeing only a few examples.
- Your learning base will remain of a very reasonable size since it uses the most recent data as a priority and you will fall into the category of responsible leaders by no longer spending fortunes in energy, bandwidth and storage to host data and process models that will only be of use to you to understand your visitor when he connects to your site and to personalize his journey.
- You are no longer subject to this “black box” of statistical modeling where, when a rule is applied – for personalization – product for example – the system is not able to explain to you why other than by “because the visitor belongs to the segment to which this rule must be applied. There the system can show you 1/ how the situation was identical 2/ what the visitors did in this identical situation.
For Scott Jones: “It will allow the next generation of AI to overcome the fragile nature of today's machine learning systems, which work well in the majority of cases but can fail spectacularly when confronted with to a case that does not fit their training model. A tragic example is the recent death of a pedestrian in Tempe, Arizona who was hit by an Uber self-driving car after the on-board AI failed to identify the pedestrian in time to take preventative action. ”
JL Bernard / Netwave CEO