A majority of machine learning models process static images. Among them, some models using Spiking Neural Networks allow for an efficient coding of the image using the sparsest representation. However, these models do not consider the predictive dynamics of the visual stream of information. Indeed, this is made difficult by the entanglement of the prediction of the features in the images with their trajectory in time. In this internship, we will build on previous work on dynamic predictive processes to propose a novel, simple and robust implementation using a Spiking Neural Network. In particular, we will use a circuit similar to an elementary micro-column to perform a form of Logistic Regression on a dynamical variable. These micro-columns will interact to provide with the best representation of local visual features, such as the orientation of a contour, from the noisy sensory input. Such a model can be potentially extended to model the interaction of different macrocolumns representing different visual positions. This will be first tested on synthetic data to validate the mathematical formulation of both the forward and backward models. This will be then applied on natural images but also on existing electrophysiological data to ponder whether the temporal dimension adds a further understanding in the observed neural activity.