The paper introduces a transition from the traditional train-recognize paradigm into a leamingbased active agent paradigm for the object recognition task. The role of different learning techniques is pointed out. We justify that a progress in the integration of learning and vision is in the development of new paradigms rather than in a simple transition of developed learning programs into the vision domain. The development of such paradigms provides an opportunity to build systems capable to learn and behave in an active manner even after the initial training is finished. We suggest that learning-based vision systems should be capable to run learning processes within the close-loop with the recognition processes; i.e., running in a "never stop learning" mode. In this case, the traditional direct distinction between the training and recognition phases tends to disappear. We indicate that the learning technology will have a major influence in the development of such autonomous active behaviors for robust object recognition systems. Some of example issues are discussed in terms of the face identification problem.