Artificial Intelligence (AI) is at the heart of the debate. Many are enthusiastic, others warn, such as Elon Musk, Bill Gates or Stephen Hawking…
AI raises many questions: What is the state of the art?
How is it defined? What are the next steps and finally what latent changes will impact our society?
This article takes up the reflections we have had in the think tank we have created, where we look at how we can improve our future by being the technological and ethical actors. Artificial intelligence has given rise to many debates, which I try to summaries below.
Where are we in terms of artificial intelligence?
In this first part, let us look at some examples that present themselves as artificial intelligence. This is not a definition, and we will see later that defining it is a vast and complex subject.
Today, we do not consider that we have reached any threshold of artificial intelligence, and these are just some basic first steps that are described below.
How can we not start with Watson? Watson is the Artificial Intelligence developed by IBM and which revealed itself in 2011 after ten years of work, by becoming the best player in the world at Jeopardy! the American general knowledge game whose answers are formulated in questions. Watson had to learn to master huge databases but also the subtleties of language.
Watson winning Jeopardy!
After this success, IBM began marketing Watson in finance, molecular biology, medicine and diagnostics. According to IBM, 75 industries now use Watson as a problem-solving tool.
How did Watson win Jeopardy! By learning from experience. To summaries, Watson is an AI based on coded rules and other mechanisms such as information retrieval. Then Watson uses Machine Learning to combine these values, assign a score and ranking to them and see if these scenarios provide the right answer.
In a few steps this results in:
- Indexing databases as Google would
- At the time of the question, search-like queries
- Several searches results
- The best 100 results are then scored and ranked
- Watson chooses one
- If it works, Watson retains the indexing, and the type of search that led to the correct answer
Watson is an example of Machine Learning. Researchers are multiplying their efforts to improve reasoning and learning systems. Deep Learning is one of the best known. It consists of trying to model an understanding that we have of the method of learning on the brain: several layers of neural networks (Recurrent Neural Networks or RNNs) that repeat the same task in order to draw conclusions, until they excel. These layers can have different functions. Deep learning techniques can automate this part of the process: detecting shapes, curves, assembling shapes, understanding objects…
The results are often both quite impressive and yet quite poor. Let me explain very quickly (a few minutes, hours), a system can be trained to recognize a picture, a dog, a baby, a shoe. But the error rate is still very high. An 8-year-old child can do the same or even better because he will never make a mistake.
Neural Networks were invented in the 1950s on the assumption that we don’t understand how the brain works, but we know that it has this kind of network. So if we simulate this type of network, maybe intelligence will emerge.
The RNN will test itself and each of the responses it produces. In short, the machine teaches itself the tasks necessary to move forward, having several layers of thinking that interact with each other.
Deep Learning was brought into the public eye with Google’s purchase of DeepMind for some $600m in 2013. The purchase followed an exercise carried out by DeepMind which consisted of becoming the best player in the world of Atari games without knowing the rules, or how the cursor worked etc. All the machines knew was that it had to maximize the number of points.
And on some of the games, in 6 hours, it had done the trick and exploded all expectations. This was obviously not the case on all the games and the learning function was in a known context: games with the same initial conditions. The scope of the work is still huge, but it is in some ways spectacular.
Demis Hassabis, Deepmind CEO presenting Atari results
And Google is not the only one interested in this. In 2014, Facebook introduced “DeepFace” which recognizes human faces at 97% (even if these are slightly hidden). We are almost at the level of a human ability on this very specific case. Yann Le Cun’s team at Facebook is working on a number of subjects to understand what information (post, news, image, etc.) is about and what it ultimately means. Microsoft and Baidu are also on the brink.
But today, machines are not yet capable of making structured and hierarchical decisions. Moreover, they are becoming extremely strong on a vertical (a specific subject) but do not have a global view.
The machine does not yet “understand” what it is doing and what it should optimize in all cases. And that is why DeepMind has not excelled in all games. Once the machine understands the method of solving complex problems on its own, a huge step in Artificial Intelligence will have been taken.
The “machines” are able to solve very quickly things that are very difficult for humans: complex equations, processing huge volumes of data… These are mainly analytical skills (calculation, data retrieval…) that are difficult for humans and appeared very recently in history.
At the same time, they are still weak in childlike tasks: recognizing an object, decoding a conversation, recognizing a face, etc. These are intuitive skills that range from recognizing emotions and places to grasping an object, etc. They have been with us for millions of years.
Computer systems are good at analytical tasks but have no intuition. It is easier to make an AI that beats the best human chess player than to make an AI that learns to walk. This is called Moravec’s paradox.
And this raises the real question of Artificial Intelligence: should it solve all human tasks and far exceed them?
To date, we don’t have Artificial Intelligence in the broadest sense, or any idea of how we’ll get there. However, we see many building blocks developing, each providing some kind of reasoning or problem solving. We will come back to this later.
Would you like to get more information about Artificial Intelligence? Contact us today.