Supervised Learning: supervised learning explained

THE Supervised Learningor supervised learning, consists in causing an AI model to make predictions or classifications for new data unknown using labeled data.

What is the Supervised Learning?

Automatic learning (Machine learning) should allow computers to recognize models and learn rules. Rather than reacting to the entry of a human user, the machines must be able to Make decisions independently Based on the rules they have learned. Algorithms can for example learn to correctly identify spam or understand the content of an image. Developers and scientists use different methods to train them. The most frequently used method is certainly the Supervised Learningthat is to say supervised learning.

In supervised apprenticeship, algorithms are driven from annotated data provided by developers, which will serve as a basis for training. Therefore, the result is already known. The task of algorithms consists only of Identify models ; Note why such information is classified in category A and such other in category B.

The supervised learning is therefore used for algorithms to classify data, whether textual, visual or audio. Furthermore, the regression problems are one of the areas of predilection application for supervised learning. In this context, algorithms must be able to carry out predictions, for example on the evolution of prices or on the increase in customers.

An intermediate approach, called Semi-Supervised learningcombines labeled and not labeled data. In this model, only part of the training data is annotated, while the rest is analyzed by the algorithm, which learns to classify the elements independently. A common example is facial recognition on social networks: after having manually identified some faces, the algorithm is able to recognize others automatically.

Supervised learning explained through an example

Suppose we wanted to train an algorithm to distinguish cat photos from dogs. The developers prepare for this purpose a very large set of data including photos already with a label, that is to say already belonging to a category. We could then imagine three groups: dogs, cats and others. It is essential that the data set shows the greatest possible diversity. To put it simply, if the training have only black cat photos, the algorithm will be the principle that all cats have a black coat. The data set must reflect the range of variants as best as possible.

During training, the algorithm first receives the raw data, makes a first classification and compares its predictions to the expected results, previously defined by the developers. He Then adjust its settings Depending on the gap between his predictions and the correct values, gradually refining his model. Learning continues until the predictions of algorithm reach a satisfactory level of precision.

Advantages and disadvantages of the Supervised Machine Learning

The choice of method largely depends on the tasks that the algorithm will have to accomplish. For the problems of regression and classificationsupervised learning is often privileged in relation to other methods. It makes it possible to train algorithms in a targeted manner so that they are perfectly suited to their field of application. This approach offers total control over the training data game, but requires a rigorous annotation and configuration work. To guarantee an efficient model, the data sample must be sufficiently varied and representative. Supervised learning therefore requires considerable efforts on the part of developers, because each data must be carefully labeled.

Although the effort is relatively important, it allows you to quite easily understand what is going on. While in not supervised learning, many things remain obscure since algorithms work without having real instructions, in supervised learning, the developer knows what the machine is doing. But this can also be a drawback: supervised algorithms are limited to the models they learned And cannot generate new knowledge outside this framework.

Differences with the Unsupervised and the Semi-Supervised Learning

In addition to the supervised learning, there is also theUnsupervised Learning (not supervised learning) and the Semi-Supervised Learning (semi-underwater learning). In what follows, we respectively approach the differences between these two methods and supervised learning.

Supervised vs unsupervised learning

While the Supervised Learning uses sets of data which we know both the inputs and the outputs, the Unsupervised Learning analysis only inputs without exit exit indication. Therefore, it aims to Discover unknown models or structures in data and is suitable for other types of tasks than supervised learning, such as cluster (Grouping of data points without classifying them in categories).

Given that the learning sets are not labeled in the case of the Unsupervised Learning, the efforts of the developers are much lower than in the case of the Supervised Learning. On the other hand, the learning process and the final result are much more opaque. It is therefore difficult to assess the performance and precision of the trained models.

Supervised vs semi-subupervised learning

A major drawback of supervised learning is the considerable time that developers must devote to data labeling. Semi-Supervised learning Use both marked and not marked data to compensate for this drawback. The model first learns from the labeled data and then improves thanks to the use of unstoppted data, recognizing models and structures.

The main advantage of the semi-subupervised learning is efficiency, given that less data must be labeled and that the process can always present a relatively high precision. It can therefore be used for classification problems similar to those of the Supervised Learning, the difference being that it will try to optimize the learning process. The complexity of the modeling and the adjustment of the balance between the labeled and not labeled data can however constitute a challenge.

Other learning methods

Supervised, not supervised and semi-supervised learning are not the only machine learning methods used to cause artificial intelligence.

THE Deep Learning is a learning process in which the models already trained learn and evolve according to their entries. These models are based on neural networks, which are inspired by the human brain.

There is also the Reinforcement Learningin French learning by strengthening, where a computer learns out of trials and errors what are the right decisions to make. It aims to establish an optimal strategy to maximize a long -term reward. An AI that learns to play a video game is an example: AI receives feedback from the training environment on each decision and thus develops game strategies.