Machine Learning Explicado: Cómo Aprenden Realmente Las Máquinas

For years, humans have built machines that follow instructions: turn on a light, turn to the left, calculate an sum. But in 1956, someone did something different. Someone built a machine that was not given explicit instructions. Instead, they allowed him to learn from the experience.

His name was Arthur Samuel and what he did was revolutionary. He created a program that allowed the ladies to play and the most amazing thing is that the more he played, the better he got.

That was the birth of what we now know as machine learning. Today, the cars that drive alone, the phones that recognize your face, the systems that predict what illness you may have by reading a radiograph, all of that is machine learning. But how does it really work? How can a machine learn?

When you were a child and learned to recognize dogs, no one gave you a list of instructions. They didn't tell you: "If he has triangular ears and hairy tail, he's a dog." Instead, you saw many different dogs. And eventually, your brain recognized the pattern. Machine learning is exactly the same, but with machines.

Formally, Machine Learning is a branch of artificial intelligence that allows computers to learn data without being explicitly programmed for each task. But how does it work? In traditional programming, you say exactly what to do. If the number is greater than 10, print "it's big". If the temperature is lower than zero, warn of the cold.

But in Machine Learning you give examples, many examples, and the computer discovers the rules by itself. It's like the difference between a traditional cooking recipe and learning to cook by watching someone do it a hundred times. Machine Learning.

It is so powerful because there are problems that are impossible to solve with explicit rules. How would you write rules to recognize faces? Each face is different. The distances between the eyes vary. The color of the skin is different. The shape of the nose is unique. But a machine learning system sees a thousand faces, learns the common patterns and can recognize new faces. Let's go back to 1956.

In IBM, an engineer named Arthur Samuel had a crazy idea: create a program that would play the ladies, but not a program where he wrote every factual movement. That would be impossible. There are more than 500,000 million of millions of legal positions in the ladies. Instead, Samuel did something ingenious: he allowed the program to play against itself.

The program began as a complete novice. He made random movements, he lost constantly, but every time he played, he learned. Samuel incorporated a mechanism where the program recorded every position he saw and if that position led to a victory or defeat.

With each game the program became more intelligent. In 1962, after playing thousands of games against himself to develop his skill, Samuel's program played against a master of ladies of real life and won.

But here is the truly revolutionary: the principles that Samuel developed, learn from experience, evaluate positions, improve through trial and error, are exactly what drive systems like ChatGPT today.

There are three different ways in which a machine can learn. Supervised learning Imagine that you want to teach a child to identify fruits. You do it by showing him an apple and saying: "This is an apple." Then a pear and you say: "This is a pear." After seeing 100 examples, the child can identify new fruits.

By feeding an algorithm with these images, the machine adjusts an internal model. Initially it will make conjectures and make mistakes, but every time it makes a mistake, it uses this feedback to correct its internal parameters and refine its precision.

With enough iterations, it will end up recognizing characteristic patterns of a fruit: shapes, colors, etc. and it will be able to distinguish them in new images. In other words, the computer learns from the experience. You teach the algorithm labeled data.

Machines only learn what they see in their training data, for good and for bad. If the data has biases or misleading patterns, the model will learn them too. A famous experiment showed that a trained model to distinguish between pictures of wolves and pictures of huskies ended up looking at something unexpected: the snow in the background.

All the photos of used wolves had snow, so the IA assumed that, if there is snow, it is a wolf. In essence, the model became a snow detector instead of an animal detector. This anecdote illustrates the importance of providing good quality data to a machine learning system.

Unsupervised learning. Now you give the child 100 objects of different types without labeling anything. You don't tell him what each thing is, you just ask him to group them. The child would look at the objects, see natural patterns and could create groups: these are metal, these are round. In a machine, the algorithm looks for patterns in data without labeling.

This approach allows the machine to "do detective" and find any regularity in the set of data. Learning by reinforcement.

In this case, the machine learns by trial and error, interacting with an environment. The correct answers are not directly told to it, but for each action it performs it receives a reward if it was a good decision or a punishment if it was bad, and its objective is to accumulate the greatest possible reward. It is similar to training a pet. Imagine a dog to whom we reward when it obeys what we ask of it and penalize when it does not.

Over time, he will prefer actions that maximize his reward.

Machine Learning is already present in countless daily apps. You wonder how Amazon knows what product to suggest, or how Spotify creates playlists you love and why Netflix knows what series you want to watch. Machine Learning algorithms analyze your habits and preferences to recommend you content to your liking, learning from your previous clicks and purchases. Your face unlocks your phone.

Facebook automatically labels you in photos. Airports scan passports. Everything is machine learning.

Modern email services use machine learning to detect spam mail automatically. They have learned to distinguish a legitimate email from a malicious one based on training. Hospitals use machine learning to detect cancer in x-rays, tumors in MRIs and predict if you will have a heart attack. Your bank uses machine learning in the background.

If you make an abnormal transaction, the system detects it based on patterns of what is "normal" for you. Cities use Machine Learning to predict congestion, suggest alternative routes and optimize traffic lights. Tesla cars constantly analyze cameras and sensors to understand the way. The system learned from millions of kilometers driven.

There is a special category of machine learning that revolutionized everything: deep learning. Deep learning uses artificial neural networks, structures inspired by how the human brain works.

Your brain has about 86 billion neurons connected to each other. Each neuron receives signals from other neurons, processes them and sends a signal to the following neurons. An artificial neural network works in a similar way. It is composed of layers of artificial neurons. Imagine that you want to identify a car in a photo. The first layer sees simple lines.

The second one combines those lines to recognize shapes. The third layer combines shapes to recognize parts. And the fourth one uses all that information and says: "It's a car!" Every time a neuron processes information, the result goes to the next layer.

the network learns by adjusting the weights of the connections between neurons. It's as if each connection had a volume that can be increased or decreased. If a connection is important for the correct decision, the volume increases. It's called deep learning because there are many layers. With enough layers, a neural network can recognize practically any pattern. For example, ChatGPT was trained with thousands of millions of phrases and human conversations

learning to predict the next word in a sentence. This way he manages to maintain dialogues and answer questions with surprising naturalness. In the same way, a generative network of images learns from millions of photos and paintings, and then he can draw something totally new that looks stylistically like what he saw. Machine learning is not the future, it is the present. And we are just starting.

Arthur Samuel showed in 1959 that a machine could learn. Almost 70 years later, what seemed science fiction is now the most fundamental technology in the world. The neural networks we saw a decade ago were considered too slow. Today they train in hours. What seems impossible today will be ordinary tomorrow. The question is not whether machine learning will transform the world. The question is: do you now understand how it works

Thank you for coming this far. If you like this content, subscribe, like it and activate the bell. Your support allows us to investigate and produce new videos. Now choose the next one on the final screen.