Minimizing binding errors using learned conjunctive features

The importance of being relevant

This review aims at an understanding of the binding process by synthesizing the extant perspectives regarding binding. It begins with a consideration of the biological explanations of binding, viz., conjunctive coding, synchrony, and reentrant mechanisms. Thereafter binding is reviewed as a psychological process guided by top-down signals. The stages and types of binding proposed by various researchers are discussed in this section. The next section introduces Working Memory (WM) as the executive directing the top-down signals. After that it is described how WM works by selecting relevant sensory input, followed by a detailed consideration of the debate regarding objects vs. features with the conclusion that relevance is the key factor determining what is processed. The next section considers other factors affecting the selection of relevant input. Then, we shift focus to describe what happens to irrelevant input - whether it is discarded at the outset or is gradually inhibited, and whether inhibition is a perceptual or post-perceptual process. The concluding section describes the process of binding as currently understood on the basis of the literature included in the review. To summarize, it appears that initially the "object" is conceptualized as an instantaneous bundle of all features. However, only relevant features of stimuli are gradually integrated to form a stable representation of the object. Concomitantly, irrelevant features are removed from the object representations. Empirical evidence suggests that the inhibition of irrelevant features occurs over time and is presumably a process within WM.

Renewing the respect for similarity

In psychology, the concept of similarity has traditionally evoked a mixture of respect, stemming from its ubiquity and intuitive appeal, and concern, due to its dependence on the framing of the problem at hand and on its context. We argue for a renewed focus on similarity as an explanatory concept, by surveying established results and new developments in the theory and methods of similarity-preserving associative lookup and dimensionality reduction-critical components of many cognitive functions, as well as of intelligent data management in computer vision. We focus in particular on the growing family of algorithms that support associative memory by performing hashing that respects local similarity, and on the uses of similarity in representing structured objects and scenes. Insofar as these similarity-based ideas and methods are useful in cognitive modeling and in AI applications, they should be included in the core conceptual toolkit of computational neuroscience. In support of this stance, the present paper (1) offers a discussion of conceptual, mathematical, computational, and empirical aspects of similarity, as applied to the problems of visual object and scene representation, recognition, and interpretation, (2) mentions some key computational problems arising in attempts to put similarity to use, along with their possible solutions, (3) briefly states a previously developed similarity-based framework for visual object representation, the Chorus of Prototypes, along with the empirical support it enjoys, (4) presents new mathematical insights into the effectiveness of this framework, derived from its relationship to locality-sensitive hashing (LSH) and to concomitant statistics, (5) introduces a new model, the Chorus of Relational Descriptors (ChoRD), that extends this framework to scene representation and interpretation, (6) describes its implementation and testing, and finally (7) suggests possible directions in which the present research program can be extended in the future.

A face feature space in the macaque temporal lobe

The ability of primates to effortlessly recognize faces has been attributed to the existence of specialized face areas. One such area, the macaque middle face patch, consists almost entirely of cells that are selective for faces, but the principles by which these cells analyze faces are unknown. We found that middle face patch neurons detect and differentiate faces using a strategy that is both part based and holistic. Cells detected distinct constellations of face parts. Furthermore, cells were tuned to the geometry of facial features. Tuning was most often ramp-shaped, with a one-to-one mapping of feature magnitude to firing rate. Tuning amplitude depended on the presence of a whole, upright face and features were interpreted according to their position in a whole, upright face. Thus, cells in the middle face patch encode axes of a face space specialized for whole, upright faces.

A model of V4 shape selectivity and invariance

Object recognition in primates is mediated by the ventral visual pathway and is classically described as a feedforward hierarchy of increasingly sophisticated representations. Neurons in macaque monkey area V4, an intermediate stage along the ventral pathway, have been shown to exhibit selectivity to complex boundary conformation and invariance to spatial translation. How could such a representation be derived from the signals in lower visual areas such as V1? We show that a quantitative model of hierarchical processing, which is part of a larger model of object recognition in the ventral pathway, provides a plausible mechanism for the translation-invariant shape representation observed in area V4. Simulated model neurons successfully reproduce V4 selectivity and invariance through a nonlinear, translation-invariant combination of locally selective subunits, suggesting that a similar transformation may occur or culminate in area V4. Specifically, this mechanism models the selectivity of individual V4 neurons to boundary conformation stimuli, exhibits the same degree of translation invariance observed in V4, and produces observed V4 population responses to bars and non-Cartesian gratings. This work provides a quantitative model of the widely described shape selectivity and invariance properties of area V4 and points toward a possible canonical mechanism operating throughout the ventral pathway.

Evidence for Holistic Representations of Ignored Images and Analytic Representations of Attended Images

According to the hybrid theory of object recognition (J. E. Hummel, 2001), ignored object images are represented holistically, and attended images are represented both holistically and analytically. This account correctly predicts patterns of visual priming as a function of translation, scale (B. J. Stankiewicz & J. E. Hummel, 2002), and left-right reflection (B. J. Stankiewicz, J. E. Hummel, & E. E. Cooper, 1998). The model also predicts that priming for attended images will generalize over configural distortions (split images), whereas priming for ignored images will not. Three experiments tested and confirmed this prediction. Split images visually primed their intact and split counterparts when they were attended but not when they were ignored, whereas intact images primed themselves whether they were attended or not. The data contribute to the growing body of evidence that 1 function of visual attention is to permit the generation of explicitly relational representations of object shape.

The temporal resolution of neural codes: does response latency have a unique role?

This article reviews the nature of the neural code in non-human primate cortex and assesses the potential for neurons to carry two or more signals simultaneously. Neurophysiological recordings from visual and motor systems indicate that the evidence for a role for precisely timed spikes relative to other spike times (ca. 1-10 ms resolution) is inconclusive. This indicates that the visual system does not carry a signal that identifies whether the responses were elicited when the stimulus was attended or not. Simulations show that the absence of such a signal reduces, but does not eliminate, the increased discrimination between stimuli that are attended compared with when the stimuli are unattended. The increased accuracy asymptotes with increased gain control, indicating limited benefit from increasing attention. The absence of a signal identifying the attentional state under which stimuli were viewed can produce the greatest discrimination between attended and unattended stimuli. Furthermore, the greatest reduction in discrimination errors occurs for a limited range of gain control, again indicating that attention effects are limited. By contrast to precisely timed patterns of spikes where the timing is relative to other spikes, response latency provides a fine temporal resolution signal (ca. 10 ms resolution) that carries information that is unavailable from coarse temporal response measures. Changes in response latency and changes in response magnitude can give rise to different predictions for the patterns of reaction times. The predictions are verified, and it is shown that the standard method for distinguishing executive and slave processes is only valid if the representations of interest, as evidenced by the neural code, are known. Overall, the data indicate that the signalling evident in neural signals is restricted to the spike count and the precise times of spikes relative to stimulus onset (response latency). These coding issues have implications for our understanding of cognitive models of attention and the roles of executive and slave systems.

Learning to recognize three-dimensional objects

A learning account for the problem of object recognition is developed within the probably approximately correct (PAC) model of learnability. The key assumption underlying this work is that objects can be recognized (or discriminated) using simple representations in terms of syntactically simple relations over the raw image. Although the potential number of these simple relations could be huge, only a few of them are actually present in each observed image, and a fairly small number of those observed are relevant to discriminating an object. We show that these properties can be exploited to yield an efficient learning approach in terms of sample and computational complexity within the PAC model. No assumptions are needed on the distribution of the observed objects, and the learning performance is quantified relative to its experience. Most important, the success of learning an object representation is naturally tied to the ability to represent it as a function of some intermediate representations extracted from the image. We evaluate this approach in a large-scale experimental study in which the SNoW learning architecture is used to learn representations for the 100 objects in the Columbia Object Image Library. Experimental results exhibit good generalization and robustness properties of the SNoW-based method relative to other approaches. SNoW's recognition rate degrades more gracefully when the training data contains fewer views, and it shows similar behavior in some preliminary experiments with partially occluded objects.