Where do concepts come from?
Speaker:
Benjamin Kuipers
Abstract:
Where do foundational concepts come from --- space, motion, places, paths, objects, actions, and so on? How can a learning agent acquire any higher-level concepts at all, starting from the "pixel level" of uninterpreted sensor and motor signals? Even if higher-level concepts are innate, they still must have been learned by the species, over evolutionary time, so we can't escape the question.
Our claim is that it is possible for developmental learning to bootstrap its way from the "blooming buzzing confusion" of pixel-level interaction with the world, to acquire higher level concepts of space, motion, places, paths, objects, actions, and so on.
This bootstrap learning can be accomplished by identifying statistical regularities in the current description of sensorimotor interaction, and positing abstract models to explain part of that regularity, factoring it away from the remaining noise. By repeating this process, the learning agent can bootstrap its way from the pixel level to a symbolic knowledge representation that describes the hypothetical outside world in terms of objects and actions.
Working with closely related simulated and physical robots, we have constructed such a developmental sequence, with most of the gaps filled in. Some steps in this sequence:
- start with an embodied agent with uninterpreted sensor and motor signals
- learn organization on the sensors and motor signals
- identify useful sensory features
- define distinctive states in terms of hill-climbing control laws
- define reliable paths linking states in terms of trajectory-following control laws
- separate and individuate objects from background
- classify objects by similar observable properties
- learn reliable actions to achieve goal states
- ability to plan to achieve goals using these higher-level actions
We believe that further research along these lines will show how qualitative abstractions can be learned that support improved predictions of the results of actions, and how this in turn can lead to learning about tools and tool-use.
Papers:Joseph Modayil and Benjamin Kuipers. 2007. Autonomous development of a grounded object ontology by a learning robot. National Conference on Artificial Intelligence (AAAI-07), 2007.
Benjamin Kuipers, Patrick Beeson, Joseph Modayil, and Jefferson Provost. 2006. Bootstrap learning of foundational representations. Connection Science 18(2): 145-158, 2006.
Jefferson Provost, Benjamin J. Kuipers and Risto Miikkulainen. 2006. Developing navigation behavior through self-organizing distinctive state abstraction. Connection Science 18(2), 2006.
Benjamin Kuipers. 2005. Consciousness: drinking from the firehose of experience. National Conference on Artificial Intelligence (AAAI-05).
Joseph Modayil and Benjamin Kuipers. 2004. Bootstrap learning for object discovery. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS-04).
Benjamin Kuipers and Patrick Beeson. 2002. Bootstrap learning for place recognition. Proceedings of the Eighteenth National Conference on Artificial Intelligence (AAAI-02).
David Pierce and Benjamin Kuipers. 1997. Map learning with uninterpreted sensors and effectors. Artificial Intelligence 92: 169-229, 1997.
