Machine Learning
Machine learning, or apprentissage automatique in French according to the CNIL, “is a field of study within artificial intelligence that aims to give machines the ability to learn from data through mathematical models.” Thus, tasks that appear simple for a human are far less so for algorithms, such as object identification.
For the machine, an object is merely a sequence of numbers, and recognizing it in every situation is impossible. Researchers have therefore found a way to work around this limitation with machine learning, which is a system trained to recognize, for example, a specific category of objects.
The goal is to enable the system to learn from data without being explicitly programmed for each task. This entire process relies on algorithms designed to identify patterns within the provided data in order to make predictions. To understand how we go from lines of Python code to a functional chatbot, it is essential to understand its learning process. The learning pipeline therefore depends on collecting relevant data to train the system, allowing it to recognize patterns that will later be applied to new data in order to generate results.
Deep Learning
This classic approach to object recognition requires building a good feature extractor, which is extremely complex and must be redesigned for each new application. It was to address this problem that deep learning came into play.
So, what is it?
It is a method whose theoretical foundation dates back to the late 1980s but that has gained popularity since 2012 with the need to train systems for multiple tasks. The concept is simple: a system is composed of several modules, each corresponding to an adjustable processing step, similar to the weights in linear classifiers, and the system is trained end-to-end. Each parameter in all modules is adjusted to bring the produced output closer to the desired output.
Training requires a calculation called backpropagation, which determines how to adjust each parameter, involving gradient computations that indicate how modifying a parameter will affect the output error. Deep learning is characterized by a multilayer network of simple components interconnected by adjustable weights, commonly referred to as a multilayer neural network. This term comes from a simplified analogy with brain neurons, where learning modifies synaptic connections between neurons. However, this model is primarily conceptual and functional and does not accurately represent brain circuits. The ability to learn and represent data hierarchically are two of deep learning’s strengths.
The first layers extract simple features, such as edges, which the subsequent layers combine into much more complex concepts making manual feature engineering unnecessary. AI, as we know it today with generative systems, requires extremely large amounts of data to generate tasks such as “speaking” or creating videos. Deep learning makes it possible to process such vast datasets and learn complex patterns thanks to multilayer neural networks.
Types of AI
There are several types of AI models with different functionalities:
- Weak AI, specialized in specific tasks such as virtual assistants. Designed to excel in limited domains, they require extensive programming to be effective.
- Strong AI, capable of performing any human cognitive task. This type of AI is still under development. Its key advantage lies in its ability to learn, enabling it to handle complex tasks that, for now, only humans can accomplish.
- Superintelligence, which refers to capacities surpassing humans in science or social interaction, but which remains purely theoretical today.
These types of AI represent a progression toward increasingly sophisticated systems whose ultimate goal is to replicate human intelligence. In this document, we will mainly focus on weak AI in order to stay grounded in practice rather than theory.
Sources :
https://www.maddyness.com/2024/02/22/ia-machine-et-deep-learning-quelles-differences/
https://www.riseup.ai/fr/blog/adaptive-learning-machine-learning
https://www.escadrille.org/introduction-intelligence-artificielle/
Juriste en droit de l’économie numérique, spécialisé en droit de l’Union européenne et en fiscalité des entreprises. J’ai des compétences en gestion de personnel, planification stratégique, gestion des ressources humaines et optimisation des processus logistiques. Grâce à mon expertise polyvalente, je navigue efficacement dans des environnements complexes pour fournir des solutions juridiques et opérationnelles adaptées aux besoins des organisations modernes.
