k-th Order Intelligences: Learning To Learn To Do

抄録

We propose a novel classification of intelligence based on distinguishing model exploitation from model exploration in order to improve our general understanding of intelligence and its limitations. For this purpose, we define computational problems by traditional function execution, which implicitly hold the model of the problem to solve, and learning problems by the meta-methods that produce computational methods. Learning problems are then assimilated to computational methods which hold implicit meta-models. The process is repeated iteratively, with each iteration named a k-th order intelligence. However, we show that the infinite sequence of classes of intelligence that emerges poses difficulties for meta-model exploration. We suggest using self-referential meta-models to break the escalation of orders, and we introduce some of the problems associated to this approach.