Zenon W. Pylyshyn

Connectionism and cognitive architecture: a critical analysis

Abstract This paper explores differences between Connectionist proposals for cognitive architecture and the sorts of models that have traditionally been assumed in cognitive science. We claim that the major distinction is that, while both Connectionist and Classical architectures postulate representational mental states, the latter but not the former are committed to a symbol-level of representation, or to a ‘language of thought’: i.e., to representational states that have combinatorial syntactic and semantic structure. Several arguments for combinatorial structure in mental representations are then reviewed. These include arguments based on the ‘systematicity’ of mental representation: i.e., on the fact that cognitive capacities always exhibit certain symmetries, so that the ability to entertain a given thought implies the ability to entertain thoughts with semantically related contents. We claim that such arguments make a powerful case that mind/brain architecture is not Connectionist at the cognitive level. We then consider the possibility that Connectionism may provide an account of the neural (or ‘abstract neurological’) structures in which Classical cognitive architecture is implemented. We survey a number of the standard arguments that have been offered in favor of Connectionism, and conclude that they are coherent only on this interpretation.

Tracking multiple independent targets: evidence for a parallel tracking mechanism.

There is considerable evidence that visual attention is concentrated at a single locus in the visual field, and that this locus can be moved independent of eye movements. Two studies are reported which suggest that, while certain aspects of attention require that locations be scanned serially, at least one operation may be carried out in parallel across several independent loci in the visual field. That is the operation of indexing features and tracking their identity. The studies show that: (a) subjects are able to track a subset of up to 5 objects in a field of 10 identical randomly-moving objects in order to distinguish a change in a target from a change in a distractor; and (b) when the speed and distance parameters of the display are designed so that, on the basis of some very conservative assumptions about the speed of attention movement and encoding times, the predicted performance of a serial scanning and updating algorithm would not exceed about 40% accuracy, subjects still manage to do the task with 87% accuracy. These findings are discussed in relation to an earlier, and independently motivated model of feature-binding--called the FINST model--which posits a primitive identity maintenance mechanism that indexes and tracks a limited number of visual objects in parallel. These indexes are hypothesized to serve the function of binding visual features prior to subsequent pattern recognition.

Computation and cognition: Issues in the foundations of cognitive science.

The computational view of mind rests on certain intuitions regarding the fundamental similarity between computation and cognition. We examine some of these intuitions and suggest that they derive from the fact that computers and human organisms are both physical systems whose behavior is correctly described as being governed by rules acting on symbolic representations. Some of the implications of this view are discussed. It is suggested that a fundamental hypothesis of this approach (the “proprietary vocabulary hypothesis”) is that there is a natural domain of human functioning (roughly what we intuitively associate with perceiving, reasoning, and acting) that can be addressed exclusively in terms of a formal symbolic or algorithmic vocabulary or level of analysis. Much of the paper elaborates various conditions that need to be met if a literal view of mental activity as computation is to serve as the basis for explanatory theories. The coherence of such a view depends on there being a principled distinction between functions whose explanation requires that we posit internal representations and those that we can appropriately describe as merely instantiating causal physical or biological laws. In this paper the distinction is empirically grounded in a methodological criterion called the “cognitive impenetrability condition.” Functions are said to be cognitively impenetrable if they cannot be influenced by such purely cognitive factors as goals, beliefs, inferences, tacit knowledge, and so on. Such a criterion makes it possible to empirically separate the fixed capacities of mind (called its “functional architecture”) from the particular representations and algorithms used on specific occasions. In order for computational theories to avoid being ad hoc, they must deal effectively with the “degrees of freedom” problem by constraining the extent to which they can be arbitrarily adjusted post hoc to fit some particular set of observations. This in turn requires that the fixed architectural function and the algorithms be independently validated. It is argued that the architectural assumptions implicit in many contemporary models run afoul of the cognitive impenetrability condition, since the required fixed functions are demonstrably sensitive to tacit knowledge and goals. The paper concludes with some tactical suggestions for the development of computational cognitive theories.

Tracking Multiple Items Through Occlusion: Clues to Visual Objecthood

In three experiments, subjects attempted to track multiple items as they moved independently and unpredictably about a display. Performance was not impaired when the items were briefly (but completely) occluded at various times during their motion, suggesting that occlusion is taken into account when computing enduring perceptual objecthood. Unimpaired performance required the presence of accretion and deletion cues along fixed contours at the occluding boundaries. Performance was impaired when items were present on the visual field at the same times and to the same degrees as in the occlusion conditions, but disappeared and reappeared in ways which did not implicate the presence of occluding surfaces (e.g., by imploding and exploding into and out of existence instead of accreting and deleting along a fixed contour). Unimpaired performance did not require visible occluders (i.e., Michottes tunnel effect) or globally consistent occluder positions. We discuss implications of these results for theories of objecthood in visual attention.

What enumeration studies can show us about spatial attention: evidence for limited capacity preattentive processing

Subitizing, the enumeration of 1-4 items, is rapid (40-120 ms/item) and accurate. Counting, the enumeration of 5 items or more, is slow (250-350 ms/item) and error-prone. Why are small numbers of items enumerated differently from large numbers of items? It is suggested that subitizing relies on a preattentive mechanism. Ss could subitize heterogeneously sized multicontour items but not concentric multicontour items, which require attentional processing because preattentive gestalt processes misgroup contours from different items to form units. Similarly, Ss could subitize target items among distractors but only if the targets and distractors differed by a feature, a property derived through preattentive analysis. Thus, subitizing must rely on a mechanism that can handle a few items at once, which operates before attention but after preattentive operations of feature detection and grouping.

Tracking an object through feature space.

Visual attention allows an observer to select certain visual information for specialized processing. Selection is readily apparent in ‘tracking’ tasks where even with the eyes fixed, observers can track a target as it moves among identical distractor items. In such a case, a target is distinguished by its spatial trajectory. Here we show that one can keep track of a stationary item solely on the basis of its changing appearance—specified by its trajectory along colour, orientation, and spatial frequency dimensions—even when a distractor shares the same spatial location. This ability to track through feature space bears directly on competing theories of attention, that is, on whether attention can select locations in space, features such as colour or shape, or particular visual objects composed of constellations of visual features. Our results affirm, consistent with a growing body of psychophysical and neurophysiological evidence, that attention can indeed select specific visual objects. Furthermore, feature-space tracking extends the definition of visual object to include not only items with well defined spatio-temporal trajectories, but also those with well defined featuro-temporal trajectories.

Return of the mental image: are there really pictures in the brain?

In the past decade there has been renewed interest in the study of mental imagery. Emboldened by new findings from neuroscience, many people have revived the idea that mental imagery involves a special format of thought, one that is pictorial in nature. But the evidence and the arguments that exposed deep conceptual and empirical problems in the picture theory over the past 300 years have not gone away. I argue that the new evidence from neural imaging and clinical neuropsychology does little to justify this recidivism because it does not address the format of mental images. I also discuss some reasons why the picture theory is so resistant to counterarguments and suggest ways in which non-pictorial theories might account for the apparent spatial nature of images.

The rate of “mental rotation” of images: A test of a holistic analogue hypothesis

This paper discusses the analogue-propositional distinction and argues that, given an appropriate understanding of this issue, the question of whether a particular cognitive function is analogue or not is an empirical one. As an example of how the question can be empirically investigated, the proposed analogue operation for mental rotation of images is considered. It is argued that the view that images are rotated in a holistic analogue manner should predict that rotation rate is independent of such factors as the conceptual complexity of the stimulus or of the comparison task. Two experiments are described that investigated the effects of several stimulus and task variables on the apparent rate of “mental rotation” of images in a Shepard-type task. Instead of comparing a stimulus and misoriented probe figure to determine whether they are identical (except for orientation) or mirror images, as was the case in most of previous studies, the present experiments required subjects to judge whether the misoriented probe was a subfigure of the target stimulus. The results showed that the “rotation rate” (i.e., the slope of the RT vs. angle of misorientation function) was influenced by practice, stimulus attributes, and the nature of the comparison task. In particular, when the probe was a “good” subfigure of the reference stimulus, apparent rotation rate was greater. These results are interpreted as indicating that the linear RT vs. angle relation is not due to a holistic analogue rotation of images, as had been supposed, but arises from a more articulated and piecemeal process in which analysis of the stimulus figure interacts with the comparison task.

Some puzzling findings in multiple object tracking: I. Tracking without keeping track of object identities

The task of tracking a small number (about four or five) visual targets within a larger set of identical items, each of which moves randomly and independently, has been used extensively to study object‐based attention. Analysis of this multiple object tracking (MOT) task shows that it logically entails solving the correspondence problem for each target over time, and thus that the individuality of each of the targets must be tracked. This suggests that when successfully tracking objects, observers must also keep track of them as unique individuals. Yet in the present studies we show that observers are poor at recalling the identity of successfully tracked objects (as specified by a unique identifier associated with each target, such as a number or starting location). Studies also show that the identity of targets tends to be lost when they come close together and that this tendency is greater between pairs of targets than between targets and nontargets. The significance of this finding in relation to the multiple object tracking paradigm is discussed.

Some primitive mechanisms of spatial attention.

Our approach to studying the architecture of mind has been to look for certain extremely simple mechanisms which we have good reason to suspect must exist, and to confirm these empirically. We have been concerned primarily with certain low-level mechanisms in vision which allow the visual system to simultaneously index items at multiple spatial locations, and have developed a provisional model (called the FINST model) of these mechanisms. Among the studies we have carried out to support these ideas are ones showing that subjects can track multiple independent moving targets in a field of identical distractors, and that their ability to track these targets and detect changes occurring on them does not generalize to non-targets or to items lying inside the convex polygon that they form (so that a zoom lens of attention does not fit the data). We have used a visual search paradigm to show that (serial or parallel) search can be confined to a subset of indexed items and the layout of these items is of little importance. We have also carried out a large number of studies on the phenomenon known as subitizing and have shown that subitizing occurs only when items can be preattentively individuated and in those cases location precuing has little effect, compared with when counting occurs, which suggests that subitizing may be carried out by counting active indexes rather than items in the visual field. And finally we have run studies showing that a certain motion effect which is sensitive to attention can occur at multiple precued loci. We believe that taken as a whole the evidence is most parsimoniously accounted for in terms of the hypothesis that there is an early preattentive stage in vision where a small number of salient items in the visual field are indexed and thereby made readily accessible for a variety of visual tasks.