Neural networks in artificial intelligence: how do they work?

Having talked about artificial intelligence and machine learning, it's time to tackle one of the most fascinating (and sometimes mystified) concepts in modern AI: neural networks.

Behind this name, inspired by the human brain, lies a powerful technology at the heart of many advances in computer vision, natural language processing, speech recognition and machine translation. In this article, we explain what neural networks are, how they work in practice, and why they are so important.

‍

‍

‍

‍

An origin inspired by biology

The idea of artificial neural networks dates back to the 1950s, when researchers such as Warren McCulloch and Walter Pitts devised a mathematical model of the neuron. Their aim was to reproduce, in a simplified way, the functioning of biological neurons: cells capable of receiving signals, processing them, and deciding whether or not to transmit information.

Artificial neural networks are therefore made up of units called neurons, organized in layers. Each neuron receives inputs, transforms them (via an activation function), then transmits an output to the following neurons.

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍

Neural network architecture

A classical neural network consists of three types of layers:

1. The input layer: receives raw data (e.g. pixels in an image).
2. Hidden layers: these carry out the intermediate calculations, transformations and decisions. The more layers there are, the deeper the network ( deep learning).
3. The output layer: this gives the final result (for example, "cat" or "dog").

Each connection between neurons is associated with a weight, which models the importance of the information transmitted.

‍

‍

‍

‍

‍

‍

‍

How does a network learn?

There are two main phases in learning a neural network: propagation and backpropagation.

1. Propagation (or forward pass): data is sent through the network, layer by layer, to the output.
2. Comparison with reality: we measure the error between the network's prediction and the true expected response.
3. Backpropagation: the network adjusts its weights by feeding information back from the output to the input, to reduce the error at the next attempt.

This process is repeated on thousands or even millions of examples, and hundreds or thousands of times or more on each example, until the network generalizes correctly.

‍

Activation functions: the little magic ingredient

Each neuron applies an activation function to its combined inputs. These functions introduce non-linearity, which is essential if the network is to model complex problems. The best-known are :

• Sigmoid: for outputs between 0 and 1
• ReLU (Rectified Linear Unit): max(0,x), widely used for its simplicity and efficiency
• Softmax: to generate probabilities for several classes

‍

Application examples

Neural networks are at the heart of many innovations:

• Image recognition: classify photos, detect faces, analyze MRIs
• Language processing: translate texts, generate summaries, answer questions
• Autonomous cars: interpreting the visual environment, detecting obstacles
• Video games: AIs capable of beating humans at highly complex games (e.g. AlphaGo)
• Voice assistants: speech recognition and synthesis

‍

Why is it important to understand neural networks?

Because they represent a new way of designing tools capable of perceiving, understanding and interacting with their environment. Understanding the basics means :

• A better understanding of the limits and possibilities of AI
• Develop a critical eye for the technologies we use
• Understanding ethical challenges (bias, opacity of models, etc.)

‍

What if you could manipulate a neural network yourself?

This is what AlphAI, the Learning Robots solution, allows you to do. With or without code, you can observe in real time how a neural network learns to react, to make mistakes, to improve... It's a fun, hands-on way to learn about these powerful concepts.

Here are our activities for a step-by-step introduction to neural networks:

👉 With the AlphAI robot:

• Start by editing the connections in the neural network yourself using the "Manual editing" activities (Blocked vs. Motion, Ultra-sound, Line tracking).
• Then use a simple example to see how a neural network learns from activity. Supervised learning - Line following which is very visual, on a simple neural network

• To go (much) further, discover with our TP Intruder detection the role of intermediate neural layers.

‍

👉 With the other robots, we recommend the activities associated with the first demo configurations in the "Manual Editing" and "Supervised Learning" categories.

‍