This article examines the primary functions of machine learning and proposes a structural framework for designing computer algorithms to address machine learning problems. A solution to the simplest machine learning problem using the simplest perceptron neural network has been proposed. The perceptron in this work is presented only as a trainable agent, that is, as a mathematical function that displays a characteristic description of the function value. The article also examines the rules for training the simplest neural network perceptron. Solutions are proposed for training examples of linearly separable classes for positive and negative data sets. Training examples show that the perceptron cannot separate any linearly separable sets of data, but only those that are divided by a line passing through the origin or a (n-1)-dimensional hyperplane passing through the origin. The article provides examples of the use of perceptron in real problems and the possible limitations that arise. The convergence rate of the trained algorithm is highly dependent on the initialization of the weight vector w. It is advisable not to change the vector at all, and the algorithm will converge after the first iteration. You can approach this using the heuristic initialization method.