Artificial intelligence is often mentioned, but its exact meaning remains unclear to many of us. In this article, we will decipher artificial intelligence by examining its recent advances, while addressing its limitations and best practices to fully benefit. We will also explore the essential prerequisites for embarking on an AI project.

What is artificial intelligence?

These are computer technologies capable of solving problems of high algorithmic complexity, especially tasks that humans perform through their cognitive functions.

To operate, AI requires several essential elements :

• A large amount of data representative of the problem to be solved
  
• Algorithms to process this data and create representation models
  
• Specific hardware for AI training and execution

Machine learning, a sub-branch of AI, has evolved considerably in recent years. Today, it no longer needs millions of data; a few can be enough to address different categories of problems, making machine learning much more accessible.

Machine learning demonstration

First, we built a database containing about 150 images, 90 with a face and 60 with an empty background. AI will be trained to recognize both the individual and the empty background, in this demonstration, this process will only take a few seconds. In a real application, training can take several days to several weeks.. Then the AI will be able to determine whether the person appears on the image or not.

We create a database, optimize the parameters of a model, and finally, so that AI can predict an output for new data. It is essential that the database be representative of the problem to be solved. Thus, it is more important to have quality data than to have a large amount of data to optimally train a machine learning algorithm.

Generative AI

Speaking of generative AI, GANs (generative adversarial network) are based on two components: the generator, whose objective is to deceive the discriminater, and the discriminater, which distinguishes real data from false ones. The better the discriminater identifies the real data, the more data the generator creates. That’s how data can be generated.

There are also AIs that work without annotation. This is the case for example of foundation models that are used by generative AI. To generate text, we train an algorithm to complete phrases we know. Properties that were not intended can then appear. For example, AI can learn to generate its own stories or do arithmetic, without being explicitly trained to do so.

However, there are limits to generative artificial intelligence. It cannot think like a human being and operates on the basis of a predefined database, which creates limits. She excels in general knowledge questions, but can provide incorrect answers on specific or complex topics.

Data and regulations

To train an AI, it is imperative to have actionable data. The collection of personal data first requires consent, followed by the anonymization of this data. In order to use raw data, their annotation is essential. To teach an AI to distinguish distinct elements, it must be shown various examples. For example, to segment an object precisely in an image and determine its nature, it is necessary to carefully annotate images pixel by pixel.

Attention to bias is crucial to minimize errors. These biases may arise from the company or the scenarios present in the databases.

It should be noted that the use of personal data is not unlimited. The GDPR, the European general regulation in force since May 2018, concerns the confidentiality and protection of personal data. It applies to data stored in Europe. Beware of foreign operators, such as the United States where American law prevails. In this context, anonymization or pseudonymization of data is essential.

In addition, the IA Act, scheduled to be implemented by the end of 2025, establishes that there is no good or bad technology, but that some use cases will be unacceptable and prohibited. This includes subliminal manipulation, exploitation of children or persons with psychological disabilities, widespread social notation (currently in force in China), and remote biometric identification (however allowed in exceptional cases, such as the search for a fugitive). In addition, some high-risk AIs will be subject to data governance, technical documentation, user transparency, and human control to ensure accuracy, robustness, and cybersecurity.

Use cases in companies