Introduction to AI #3: technical approach
Artificial Intelligence (AI) raises many questions. And for a good reason: it is an extremely broad area for which there are many definitions and which disrupts many aspects of society.
If we hear a lot about AI today, it is because it has an increasingly more significant place in our lives: indeed, we use tools based on AI every day. For instance, it helps us with automatic word completion on mobile phones, it allows the recommendation of press articles or information in a newsfeed or it is used to optimize the price of airplane tickets.
In the previous articles , we have traced the origins of AI back, from automatons to the birth of computers in the aftermath of the Second World War. From the 2000s, technical progress and automation supported the development of AI. In this article, we present some key technical terms which will help to clarify how AI works.
AI: a multi-level technology
First of all, AI can be divided into three levels of complexity:
- Expert systems. Those are the simplest systems, they are based on predefined rules. For example, opponents in video games are expert systems: the AI simulates another player. However, in this case, each situation has been predicted and the reaction has been programmed: this is the major limitation of expert systems. They are the part of AI which does not learn, unlike the two categories below.
- Machine learning. Those systems are a little more complex, they are based on models designed by humans but able to adapt according to the patterns observed with the data. This is the case with the recommendation engines in e-shopping websites: the AI simulates a salesman. In this case, most of the time, the algorithm is able to learn from its successes and it improves the model on which it is designed.
- Deep learning. Those systems are able to adjust their models themselves according to the data patterns. The latest example is AlphaGo, Google’s Go player: the AI simulates a player and it has managed to beat the best Go human players in the world in 2016 and 2017. Deep learning is a sub-category of machine learning.
Weak AI VS. Strong AI
After this division into complexity levels, one might be tempted to consider the terms strong AI and weak AI. Those terms are sometimes found when it comes to AI… and yet, they were not invented by computer scientists but by the philosopher John Searle! He invented the term strong AI to challenge the capabilities of AI and the fact of consedering machines as a model of human thinking. Weak AI would represent an AI which does not have any conscience, emotion, will and whose actions are very targeted. On the contrary, strong AI would have the same conscience and abilities than humans.
Today, it is impossible to create a strong AI. It even seems unlikely to succeed in creating one in the future although it is possible, for some extremely specialized tasks, to come to AIs which perform better than humans. Besides, this performance requires a much longer training, more energy and more data than human learning.
Focus on machine learning
What do we mean when we say that an algorithm (or a machine) has learned “on its own”? This terminology often refers to two sub-categories of machine learning: reinforcement learning and unsupervised learning. Machine learning can be divided into three categories:
- Supervised learning: in this case, the learning is based on complete information, which means that the data set describes an object and one or several target variables to be predicted. For example, an organ is the object described by an X-ray and the AI will try to predict whether there is or will be a tumor in the next 6 months. To do so, the X-rays were labelled with the data “yes” or “no”: this allows the algorithm to learn with a given measurement, before being used to predict the target data when it is unknown.
- Reinforcement learning: in this case, there is no “correct” answer but there is a notion of gain and loss which allows the algorithm (called agent) to identify which action is positive and which action is negative. This method is used to teach robots to walk: if they move forward, they win while if they fall, they lose. By testing different combinations of actions, they are able to find one which allows them to move forward – but which is not necessarily the best.
- Unsupervised learning: in this case, the learning data is not labelled. The objective is to understand a pattern in the data. For example, from pictures of different flowers, the AI tries to gather the pictures representing flowers of the same species.
This trilogy ends on a focus on machine learning, a technology which is part of our lives. But what about its use at Domoscio? Check our website or contact us!
 Quote from Laurence Devillers, professor of applied computer science in social science at the University of Paris-Sorbonne 4 since 2011.
Our latest articles about the topics: learning, science and Domoscio's news.