Scaling Up Grounded Representations Hierarchically

We have been studying the learning of compositional hierarchies in predictive models, an area we feel is significantly underrepresented in machine learning. The aim in learning such models is to scale up automatically from fine-grained to coarser representations by identifying frequently occurring repeated patterns, while retaining the ability to make predictions based on the statistica] regularities exhibited by these patterns. Our hierarchical learning begins with data consisting of discrete symbols and can be viewed as a method of grounding high-level concepts in terms of their lower-level parts, which are themselves grounded in raw, environmental signals by other means. This short paper discusses the relationship between hierarchy learning and the learning of low-level grounded representations and also very briefly describes one of our systems for compositional hierarchy learning. A much more detailed discussion, including an extensive literature review, can be found in (Pfleger 2000). Compositional Hierarchy Learning Many AI systems operate using representations at levels well above raw environmental data. Whether they are called states, operators, features, rules, schemata, scripts or some other name, the successes or failures of systems have often depended more critically on human choices for instantiating the representational units than on the choices of algorithms or the designs of the representational languages themselves. In order to design highly intelligent autonomous agents with flexibility and aptitude far surpassing that which has been artificially created today (Nilsson 1995; Newell 1990; Hayes-Roth 1993), a vast number of representational units spanning many levels of resolution and abstraction will be needed, and it seems evident that learning must play a role in developing these representations. We agree with (Sun 2000) that high-level representational units must be grounded in the environment. This includes the implications that both representation Copyright © 2001, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. and environmental interaction are important. We also agree that a bottom-up learning process is necessary to create the high-level units. 1 This vague notion of a continuum from low-level to high-level representations conflates two correlated measures of representational level based on two distinct types of structure. Taxonomic structure relates entities by is-a relationships, and compositional structure relates entities by part-of relationships. Machine learning and related fields have a long history of producing systems that learn taxonomic hierarchies, but the field has not made corresponding progress at building compositional hierarchies (CHs). In fact, the building CHs has only surfaced in a few areas and in each case the learning depends on specialized features of the subdiscipline, such as the top-down use of separate domain theories. We have been attempting to fill this gap by investigating bottom-up, data-driven methods for learning compositional hierarchies. Our methods are domainindependent and action-independent, and they learn in an on-line fashion. The methods operate on unbounded, stationary data (e.g., unbounded sequences) that consist of discrete symbols drawn from a finite alphabet. Learning is unsupervised, and the models are capable of arbitrary statistical inference (prediction). The learning paradigm is analogous to the standard unsupervised prediction paradigm for IID data, adapted to unbounded sequences or higher dimensional analogs. See (Pfleger 2000) for precise details on the learning paradigm. The general strategy for incrementally building CHs is to repeatedly combine, or chunk, frequently occurring patterns into higher-level aggregates, enabling the future combination of these patterns into even larger, higher-level aggregates. Thus, the stance taken here is that representations are learned by noticing and storing patterns exhibited by the environment. Once learned, CHs lead naturally to smooth integration of bottom-up and top-down processing that can mediate lateral in1We do not, however, agree that decisions about actions are the necessary type of environmental interaction for the creation of all forms of high-level concepts. See (Pfleger 2000). 56 From: AAAI Technical Report SS-01-05. Compilation copyright © 2001, AAAI (www.aaai.org). All rights reserved.