Leveraging Failures to Enhance Hierarchical Concept Learning when Training and Testing are Limited

Hierarchical concept learning constructs higher-level concepts using previously learned prerequisite concepts. We are working in an especially challenging context where only a small number of training instances for each concept are provided to the learning system. This limited instruction forces even the most skillful learner to make assumptions about the concept being taught—assumptions that can be incorrect. Given this uncertainty, multiple candidates may be proposed for the concept, each stemming from different assumptions that are consistent with the training. We present a control strategy for managing the use of hypothesized concept candidates in higher-level learning. The strategy is based on three key ideas: 1) limiting prerequisitecandidate combinatorics by operating with a single selected candidate for each concept at any time, 2) using learning failure to select a different candidate for a direct or indirect prerequisite concept, and 3) using differences observed as candidates are used to guide candidate selection. We implemented and evaluated this novel Concept Candidate Management (CCM) strategy in MABLE, an electronic student that performs bootstrapped concept learning. Using the CCM strategy, MABLE learned concepts that were not successfully learned otherwise—without any additional training or testing and without any changes to learning algorithms.