Towards open ended learning: budgets, model selection, and representation

Abstract

Biological organisms learn to recognize visual categories continuously over the course of their lifetimes. This impressive capability allows them to adapt to new circumstances as they arise, and to flexibly incorporate new object categories as they are discovered. Inspired by this capability, we seek to create artificial recognition systems that can learn in a similar fashion. We identify a number of characteristics that define this Open Ended learning capability. Open Ended learning is unsupervised : object instances need not be explicitly labeled with a category indicator during training. Learning occurs incrementally as experience ensues; there is no training period that is distinct from operation and the categorization system must operate and update itself in a timely fashion with limited computational resources. Open Ended learning systems must flexibly adapt the number of categories as new evidence is uncovered. Having identified these requirements, we develop Open Ended categorization systems based on probabilistic graphical models and study their properties. From the perspective of building practical systems, the most challenging requirement of Open Ended learning is that it must be carried out in an unsupervised fashion. We then study the question of how best to represent data items and categories in unsupervised learning algorithms in order to extend their domain of application. Finally, we conclude that continuously learning categorization systems are likely to require human intervention and supervision for some time to come, which suggests research in how best to structure machine-human interactions. We end this thesis by studying a system that reverses the typical role of human and machine in most learning systems. In Crowd Clustering , humans perform the fundamental image categorization tasks, and the machine learning system evaluates and aggregates the results of human workers.