An agent moving in a real environment perceives it by numerous noisy sensors which provide some high dimensionality data with unknown topology. In order to interact in this complex and changing environment, according to the active perception theory, the agent needs to learn the correlations between its actions and the changes they induce in the environment. In the perspective of a bio-inspired architecture for the learning of multi-modal correlations, this article focuses on the ability to forget some previously learned selectivity in a model of perceptive map which spatially codes the sensor data. This perceptive map combines the BCM (Bienenstock Cooper Munro) learning rule, which raises a selectivity to a stimulus, with the neural field (NF) theory, which provides spatial constraints to self-organize the selectivities at the map level. The introduction of an unlearning term in the BCM learning rule improves the BCM-NF coupling by providing plasticity to the self-organization.