Title: What is the Ground? Continuous Maps for Symbol Grounding
Abstract: What is the Ground? Continuous Maps for Grounding Perceptual Primitives Ian Perera ([email protected]) and James F. Allen ([email protected]) Department of Computer Science, University of Rochester, 500 Joseph C. Wilson Blvd., Rochester, NY 14627 USA Institute for Human and Machine Cognition, 40 South Alcaniz Street Pensacola, FL 32502 USA Abstract physical attributes and their combinations to learn about ob- jects it sees (Farhadi, Endres, Hoiem, & Forsyth, 2009), but would also be built around a framework allowing evaluation of claims that it has actually grounded such attributes in a perceptual model. Various types of reasoning are also supported only by a per- ceptually grounded system. Reasoning by mental simulation requires a grounding in physical form to process questions such as “Can an open umbrella fit in the trunk of a car?” 1 , which is trivial for a person to answer even though the an- swer is not explicitly stated in a knowledge base. Qualitative reasoning also draws upon non-symbolic knowledge in rea- soning over continuous spatial, temporal, or feature spaces. Rather than explicitly stating the properties and relations be- tween objects for every possible pair of objects, a grounded system can draw upon its perceptual representations to an- swer questions about relations between objects on the fly. Analysis of the Symbol Grounding Problem has typically fo- cused on the nature of symbols and how they tie to percep- tion without focusing on the actual qualities of what the sym- bols are to be grounded in. We formalize the requirements of the ground and propose a basic model of grounding perceptual primitives to regions in perceptual space that demonstrates the significance of continuous mapping and how it influences cat- egorization and conceptualization of perception. We also out- line methods to incorporate continuous grounding into compu- tational systems and the benefits of applying such constraints. Keywords: Symbol Grounding; Perception; Language; Ma- chine Learning; Topology Introduction The symbol grounding problem underlies a wide range of areas in cognitive science, including perception, philosophy of meaning, child language learning, and artificial agents – “How can the semantic interpretation of a formal symbol sys- tem be made intrinsic to the system, rather than just parasitic on the meanings in our heads?” (Harnad, 1990). The answer to this question – how do our words have meaning beyond definitions composed of other words – can guide our models of cognition and our algorithms of artificial agents, providing constraints on the grounding process and enabling capabili- ties not present in ungrounded systems. While Harnad clearly states what a symbol system is – a system of explicit representation, syntactic manipulability, semantic interpretability, and systematicity – he does not go into great detail as to what the symbols are grounded in. At- tempts to formalize the Symbol Grounding Problem focus on the symbolic aspect, but neglect to incorporate a theoretical categorization of acceptable grounds. This gap in specificity is a detriment to research in the area of symbol grounding – how do we design perceptually grounded systems without knowing what capabilities we are pursuing nor what kinds of systems have these capabilities? Previous Work Advantages of a Perceptually Grounded System To say that an ungrounded system is lacking because its knowledge is circularly-defined is not entirely satisfying if there is no appreciable performance difference between a grounded and ungrounded system. The primary advantage is that a perceptually grounded system is uniquely capable of interfacing its knowledge with perceptual systems, even be- yond current computer vision and robotics systems. A percep- tually grounded ontology, for example, could be used to rec- ognize objects that had never been seen before based on sym- bolic knowledge such as in work by Russakovsky and Fei-fei (2012). A computer vision system could not only recognize There have been a number of efforts attempting to pro- vide a theoretical representation that crosses the divide be- tween symbolic knowledge and perceptual representations. Barsalou (1999) proposed a perceptual theory of knowledge to explain the connection between symbolic knowledge and perception. At the center of the theory is the perceptual sym- bol, a neural representation that is tied to the underlying sense modality related to its symbol. A visual symbol, for example, is a record of the neural state in the visual cortex. Perceptual symbols are composed of perceptual components, which are the primitives corresponding to sensory features. In this work we consider grounding at the feature level and therefore are concerned with the perceptual components – to convey this level of grounding, we will refer to such components as per- ceptual primitives. G¨ardenfors (2004) proposed his theory of conceptual spaces as a non-symbolic framework for information pro- cessing, representing concepts as regions in spaces defined according to quality dimensions of the stimulus. His use of the term conceptual space indicates that in his model, con- cepts are identical to the regions in what we define as percep- tual space: the space where stimuli are arranged according to a particular set of qualities. We use the term perceptual space to make a distinction between the range of possible stimuli and the symbols that are associated with them – sym- bols which might be distributed according to conceptual sim- ilarity in a different space. In attempting to describe the geometrical constraints on re- gions in conceptual space, G¨ardenfors claims that in such a 1 This example originated with Lenhart Schubert.
Publication Year: 2014
Publication Date: 2014-01-01
Language: en
Type: article
Access and Citation
AI Researcher Chatbot
Get quick answers to your questions about the article from our AI researcher chatbot