Historically, the psychophysical evidence for "selective attention" originated mainly from visual search experiments. A first important distinction in the processing of information in visual search tasks is its separation in two stages. The first, early "preattentive" stage operates in parallel across the entire visual field extracting single "primitive features" without integrating them. The second "attentive" stage corresponds to the specialized integration of information from a limited part of the field at any one time, i.e. serially. So far, models based on the above mentioned two-stage processes have been able to distinguish features from conjunction search conditions based on the observed slopes of the linear relation between reaction time (i.e., search time) and the number of items in the stimulus array. We propose a neuroscience based model for visual attention that works across the visual field in parallel, but due to its intrinsic dynamics can show the two experimentally observed modes of visual attention, namely: the serial focal attention and the parallel spread of attention over space. The model demonstrates that neither explicit serial focal search nor saliency maps need to be assumed. In the present model the focus of attention is not included in the system but only emerges after convergence of the dynamical behaviour of the neural networks. Furthermore, existing models have not been able to explain the variation of slopes observed in different kinds of conjunction search modes. We hypothesize that the different slopes can be explained by assuming that selective attention is guided by an independent mechanism which corresponds to the independent search for each feature. The model consistently integrates the different neuroscience levels by considering the microscopic neurodynamical mechanism that underlies visual attention, the different brain areas of the dorsal or "where" and ventral or "what" paths of the visual cortex, and behavioural data.