Unbiased global illumination methods based on stochastical techniques provide photorealistic images. However, they are prone to noise that can only be reduced by increasing the number of processed samples. The problem of finding the number of samples that are required in order to ensure that most observers cannot perceive any noise is still an open issue. In this article, we address this problem focusing on visual perception of noise. However, rather than using known perceptual models, we investigate the use of learning approaches classically used in the field of Artificial Intelligence. Hence, we propose to use such approaches to create a model which is able to learn which image highlights perceptual noise. The learning is performed through the use of a database of examples based on experimentations of noise perception with human users. This model can then be used in any progressive stochastic global illumination method in order to find the visual convergence threshold of different parts of an input image.