Human cognition heavily relies on inductive learning, a process that the field of machine learning aims to replicate in artificial hardware/software. While connectionist learning methods have yielded great pragmatic results in this area, they still lack a model hierarchy of learning to explain their results. NeSy computing seeks to develop effective integration between connectionist and symbolic learning. As an effort to achieve this integration, NeMuS is a multi-dimensional graph structure, originally conceived with four spaces of codified elements of first-order logic, that learns patterns of refutation and performs inductive clausal reasoning to induce hypotheses that explain examples non-previously specified in a background knowledge. Recently, there was an experiment in which a connected background knowledge was trained using SOM to generate similarity of concepts according to their attributes, and their respective position within the concepts. In this experiment, induction was performed by human analysis on the organizational map of concepts. In this work, we seek a suitable method to generate neighbourhood patterns to be used for inductive learning and reasoning in order to reduce the search space of hypotheses. Additionally, we define a language bias able to handle predicate invention, to guide the process of generating such hypotheses.
|
