Recognition of self-controllable body parts via online Multiple Instance Learning
On-line adaptation is an essential capability for artificial cognitive agents operating in the real world. As the scientific community devotes growing attention to the development of such systems, there is the need to shift from well-established batch algorithms to learning techniques in which data is acquired autonomously and employed on-line. Autonomous learning implies that data is collected automatically without human supervision. Learning in real world applications is therefore hampered by a unfavorable tradeoff between the accuracy of the training examples and their availability. One way to alleviate this problem in classification contexts is to adopt a semi-supervised learning paradigm such as Multiple Instance Learning (MIL). In this framework, training examples come in the form of positive or negative "Bags" of instances and the system needs to learn which of such instances are responsible for the positive or negative label of the bag. We present an online MIL algorithm which exploits an online variant of Adaboost combined with a family of weak hypotheses specifically developed for the purpose of multiple instance learning. A practical application of the presented framework is provided in the visual learning domain: a robotic system learns to recognize its own controllable body parts (in particular the hand) via self observation.