Towards combining commonsense reasoning and knowledge acquisition to guide deep learning

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