Beyond Data: Efficient Knowledge-guided Learning for Sparse and Structured Domains

Abstract

The field of AI has made great advances in recent years. Most of these advances have focused on leveraging more data and finding new architectures to improve system performance. However, collecting data can lead to exorbitant costs. This is especially the case for structured domains where the data conforms to some standardized format (like tabular data, relational databases, etc.). In structured domains, an expert might be required to collect and organize data; necessitating time and effort. Further, learning explicitly from data is neither sufficient nor favorable. Enormous data can cause concerns for safety, lack of fairness, and a substantial carbon footprint. So looking beyond learning from data, this dissertation focuses on finding principled ways to leverage rich human knowledge for sparse and structured domains to guide the learning procedure. In particular, this dissertation looks at four challenges that arise when models are learned in structured domains and propose to tackle them using explicit human knowledge. First, we consider the challenge of learning from sparse and noisy data in the successful gradient boosting framework and propose to use domain-specific trend information to improve prediction. Second, we consider the challenge of learning to generalize across multiple tasks and objects in sequential decision making. We address this challenge by proposing a framework that takes inspiration from human’s ability to generalize by identifying compositionality and generating abstract representations. Third, we consider the challenging task of human-machine collaborative problem solving and propose a framework that uses natural language communication for effective bi-directional interaction. Finally, the fourth challenge we consider is the problem of a large hypothesis space when dealing with domains with heterogeneous objects. We identify the lack of a language bias—typed object representations—in recent neurosymbolic architectures and devise an approach to incorporate the bias. In this dissertation, we demonstrate various ways to incorporate domain-specific knowledge from humans in training AI systems. We conclude that using domain knowledge not only reduces the sample complexity but also improves the performance and generalization abilities of the model.