Algorithms based on deep network models are being used for many pattern recognition and decision-making tasks in robotics and AI. Training these models requires a large labeled dataset and considerable computational resources, which are not readily available in many domains. Also, it is difficult to explore the internal representations and reasoning mechanisms of these models. As a step towards addressing the underlying knowledge representation, reasoning, and learning challenges, the architecture described in this paper draws inspiration from research in cognitive systems. As a motivating example, we consider an assistive robot trying to reduce clutter in any given scene by reasoning about the occlusion of objects and stability of object configurations in an image of the scene. In this context, our architecture incrementally learns and revises a grounding of the spatial relations between objects and uses this grounding to extract spatial information from input images. Non-monotonic logical reasoning with this information and incomplete commonsense domain knowledge is used to make decisions about stability and occlusion. For images that cannot be processed by such reasoning, regions relevant to the tasks at hand are automatically identified and used to train deep network models to make the desired decisions. Image regions used to train the deep networks are also used to incrementally acquire previously unknown state constraints that are merged with the existing knowledge for subsequent reasoning. Experimental evaluation performed using simulated and real-world images indicates that in comparison with baselines based just on deep networks, our architecture improves reliability of decision making and reduces the effort involved in training data-driven deep network models.