Towards a Map for Incremental Learning in the Limit from Positive and Negative Information

In order to model an efficient learning paradigm, iterative learning algorithms access data one by one, updating the current hypothesis without regress to past data. Prior research investigating the impact of additional requirements on iterative learners left many questions open, especially in learning from informant, where the input is binary labeled. We first compare learning from positive information (text) with learning from informant. We provide different concept classes learnable from text but not by an iterative learner from informant. Further, we show that totality restricts iterative learning from informant. Towards a map of iterative learning from informant, we prove that strongly non-U-shaped learning is restrictive and that iterative learners from informant can be assumed canny for a wide range of learning criteria. Finally, we compare two syntactic learning requirements.