The olfactory system, and in particular its first relay center, the olfactory bulb (OB) is a ``working horse'' for investigating the structural conditions and functional significance of chaotic information processing in neural networks. Oscillatory activities were studied in the framework of a model based on the connections among oscillators formed by pairs of an excitatory mitral and an inhibotory granule cell. In the present article, the extended version of this model, also taking into account the lateral connections in the layer of mitral cells, is systematically studied. Because of the lack of firm data on the physiological nature of the lateral interactions in the mitral layer, the potential effects of both inhibition and excitation on the qualitative dynamic behavior are studied. The strengths of the lateral connections are taken to be constant over time, and can be varied through a control parameter. The temporal character of the sensory input (``sniff cycle'') is neglected in this series of simulation experiments. In another series of simulations we explicitly take into account the modifiability of certain synaptic strengths. More precisely, the excitatory mitral-mitral interactions are modified by a slow process prescribed by a learning rule. A specific rule has been designed to describe learning in and recall from the olfactory bulb. The learning rule consists of a nonlinear decay term, a Hebbian term, and a selective decrease term. The increase of synaptic strengths due to learning implies transitions between different qualitative regimes. A transition from an oscillatory to a chaotic regime is demonstrated. Taking into account the explicit time-dependence of the sensory input, the mathematical model is no more an autonomous system of ordinary differential equations; instead, it is a nonautonomous system. A nonautonomous system does not even have an attractor in the general case; therefore, the concept of ``computation with attractor'' cannot automatically be applied for modeling associative memory phenomena. Continuous-time learning also causes difficulties since it distorts or destroys the ``innate categories'' defined by the structure and parameters of an autonomous system. Another problem is that traditional memory models deal with the classification of static (and not time-dependent) inputs. In this article both theoretical remarks on the classification problem are made, and the learning/memory problem in the OB is studied, which is considered as a case study of information processing with continuous learning and nonautonomous dynamic systems.