Commentary on Hurford, James R.

Abstract: 90 words
Main Text: 1050 words
References: 100 words
Total Text: 1140 words

Pre-linguistic agents will form only ego-centric representations

Michael L. Anderson
Institute for Advanced Computer Studies
University of Maryland
4447 A.V. Williams Building
College Park, MD 20742

Tim Oates
Department of Computer Science and Electrical Engineering
University of Maryland Baltimore County
1000 Hilltop Circle
ECS Building Room 214
Baltimore, MD 21250


The representations formed by the ventral and dorsal streams of a pre-linguistic agent will tend to be too qualitatively similar to support the distinct roles required by PREDICATE(x) structure. We suggest that the attachment of qualities to objects is not a product of the combination of these separate processing streams, but is instead a part of the processing required in each. In addition, we suggest that the formation of objective predicates is inextricably bound up with the emergence of language itself, and so cannot be cleanly identified with any pre-linguistic cognitive capacities.

In his search for the neural basis of the simple logical structure PREDICATE(x), Hurford focuses on the basic cognitive capacities of advanced, but pre-linguistic agents--for he follows (Batali 2002) in believing that language can be developed out of such simple logical structures. Not having any such agents to study directly, Hurford instead considers the capacities of higher primates, abstracting away from their linguistic abilities to uncover two basic perceptual processing systems--the dorsal and ventral streams (henceforth DS and VS)--which he suggests provide the basic components of PREDICATE(x) structure. However, he does not carry this thought experiment through consistently, and, failing to consider matters from the perspective of such a pre-linguistic agent, he is lead to interpret the information delivered by these perceptual processing streams in post-linguistic terms, thereby in a subtle way assuming what he is trying to prove. When things are instead considered from the standpoint of such an agent, in no longer looks as if the DS and VS provide attractive candidates for the separate components of PREDICATE(x) structure.

Hurford rightly emphasizes that the deliverances of the DS--the "where" pathway that provides information about the location and size and shape of an object--are cast in "ego-centric" terms. The DS is a specialized perceptual processing system that represents information in a form optimized for calculating and directing motor responses aimed at an object in virtue of its location, orientation, and spatial extent. This information is used to guide such things as the orientation of sense organs for optimal perception, perceptual tracking, reaching, and grasping. Thus, the natural way to characterize what the agent knows in virtue of DS representations (what the information means to the agent) is in terms of ego-centric spatial coordinates: where it is in relation to the agent, and what might be done to get the agent-object relation into a preferred state. One might say that the DS places objects in an ego-centric visuomotor space, or an ego-centric action field, and the object is thereby presented to the agent in these terms.

This characterization of the function of the DS is largely in accord with Hurford's--but what drives Hurford's account is the supposed contrast between the ego-centric "motor-oriented" information given by the DS, with the "cognitive"--and therefore in some sense more objective--information said to be delivered by the VS. But in the individual, pre-linguistic, and thus (one might say) functionally solipsistic agents Hurford describes, this contrast is untenable. Like the DS, the VS is a specialized perceptual processing system, but in this case optimized for representing information about the look rather than the location of an object. Just as with the DS, the VS representations are used to select and direct appropriate motor responses; and just as with the DS, the most natural way to characterize what the agent knows in representing this information is: what the object means to it--to its utility, goals, survival--and what it might appropriately do in response. VS representations likewise place the object in a visuomotor action field, useful for calculating responses based on the differences between individual objects, as opposed to individual locations. To instead imagine that encountered objects are represented in terms of objective features or abstract qualities is to import into the VS the kind of representational scheme appropriate for language and logic, but of no use to agents whose primary concerns are individual and behavioral, and not communal or communicative.

So what we in fact have in the case of the individual agents Hurford describes are two specialized processing mechanisms which, although optimized for representing different aspects of perceptual information, are both nevertheless engaged in interpreting that information in conceptual terms suited to the selection and direction of appropriate motor responses. It does not seem that the products of these two visuomotor control systems lend themselves to natural combination in the form Hurford needs. Indeed, there seem to be substantial gaps between what these pathways deliver--the ego-centrically presented features of objects--and PREDICATE(x) structure. We describe two of these gaps, below, and identify possible approaches to bridging them.

Consider first the fact that the representations formed by the VS and DS are ego-centric. Language is useful only to the extent that it enables agents to share meaning with one another, but for meaning to be shared it must be objective, not subjective or idiosyncratic. An isolated agent capable of forming pre-linguistic concepts, or predicates, from the information delivered by its VS will form just those concepts that help this one agent survive in whatever environment it finds itself. These concepts will be cast within a single, agent-centered frame of reference, meaningful only from its own individual perspective. But suppose this agent discovers others of its kind. Through repeated interactions with these new agents, objective features of the world--those features commonly available and salient to others--can be identified and thus used to form the concepts (predicates) that serve as the semantic basis of language. The computational model of language evolution developed by Luc Steels (Steels 1997) leverages this idea of repeated interactions to separate the objective from the subjective and thereby evolve a stable, shared lexicon. Genuinely objective predicates, and the representational schemes which support them, arise only as the result of the formation of such shared, stable, inter-subjective representation systems.

Second, it's clear that in PREDICATE(x), the two components--PREDICATE() and x--are qualitatively different. But if both processing streams are delivering the ego-centrically presented features of objects, then neither stream seems properly specialized for producing objects over properties. Rather, if both pathways can be said return information about objects--such as the distance to an object in the case of the dorsal pathway, or the color of an object in the case of the ventral pathway--then it seems that both pathways must have access to (or contain) a neural mechanism that individuates objects in the visual field, making it possible to bind deictic markers and extract information about, for example, distances or colors. Given this mechanism, the dorsal pathway could produce representations to underlie predicates like REACHABLE(x), and the ventral pathway could produce representations to underlie predicates like RED(x). Indeed, it seems that the data presented in (Goodale et al. 1994) support this view, for the patient R.V. was able to discriminate objects despite apparent damage to DS processing.


Batali, J. (2002). The negotiation and acquisition of recursive grammars as a result of competition among exemplars. In T.Briscoe (Ed.), Linguistic evolution through language acquisition: Formal and computational models. Cambridge University Press.

Goodale, M.A., J.P. Meenan, H.H. Bülthoff, D.A. Nicolle, K.J. Murphy, and C.I. Racicot (1994). Separate neural pathways for the visual analysis of object shape in perception and prehension. Current Biology 4(7), 604-10.

Steels, L. (1997). Constructing and sharing perceptual distinctions. Coupling of meaning creation and naming games in software experiments. In: van Someren, M. and G. Widmer (Eds.), Proceedings of the European conference on machine learning. Springer-Verlag, Berlin.