Learning Symbolic Latent Representations for Relational Data

This research project addresses the problem of statistical predicate invention in machine learning. Machine learning is a branch of artificial intelligence that focuses on the development of algorithms whose performance improves with experience. Such algorithms learn from examples, instead of being explicitly programmed for the task and are often used to solve the tasks we do not have a solution for. More precisely, this project will follow the framework of statistical relational learning - a branch of machine learning that combines machine learning with logic. Predicate invention represents the discovery of new concepts, properties and relations from data, if they exist, and expressed in terms of the observedones, using statistical techniques to guide the process and explicitly representing the uncertainty in the discovered predicates. It is considered to be one of the key problems in statistical relational learning as it allows to improve a language in which data is represented. The representation is the key factor causing learning to be easy or difficult. It is, therefore, advantageous to first learn a good representation, then learn predictive models using this representation. The objectiveof this research is to (1) develop more powerful methods for predicate invention and investigate its relation to (2) structure learning and (3) transfer learning. This will advance the state of the art in machine learning in several ways: it will allow more accurate and compact models, and learning layered logical programs. The problem of structure learning refers to the problem of identifying (in)dependenciesbetween variables or predicates. We speculate that the structural (in)dependencies could be identified through predicate invention. Finally, wewould like to investigate its connections to transfer learning - the problem of transferring knowledge gained while solving one problem to a different, but related one. The approach chosen for this researchfollows a recently proposed perspective on predicate invention based onunsupervised learning. We will address the problem from a perspective of deep learning which offers a wide range of unsupervised learning methods that haven not been tested in this context yet. Not only it provides a wide range of methods to test, deep learning also resembles the idea of predicate invention. Deep learning is an approach to learning where the input is not directly mapped to the target output, but multiple intermediate representations are learned in an unsupervised manner. These multiple intermediate layers represent the abstractions of previous layers have a lot in common with predicate invention. Motivated by the indicatedsimilarity between deep learning and predicate invention, and a successof deep learning, the main question we want to address is how methods from deep learning could be used for the task of predicate invention.

Publication year:2018

Accessibility:Open