This paper proposes a novel approach to constructing a hierarchical representation of visual input that aims to enable recognition and detection of a large number of object categories. Inspired by the principles of efficient indexing (bottom-up), robust matching (top-down), and ideas of compositionality, our approach learns a hierarchy of spatially flexible compositions, i.e. parts, in an unsupervised, statistics-driven manner. Starting with simple, frequent features, we learn the statistically most significant compositions (parts composed of parts), which consequently define the next layer. Parts are learned sequentially, layer after layer, optimally adjusting to the visual data. Lower layers are learned in a category-independent way to obtain complex, yet sharable visual building blocks, which is a crucial step towards a scalable representation. Higher layers of the hierarchy, on the other hand, are constructed by using specific categories, achieving a category representation with a small number of highly generalizable parts that gained their structural flexibility through composition within the hierarchy. Built in this way, new categories can be efficiently and continuously added to the system by adding a small number of parts only in the higher layers. The approach is demonstrated on a large collection of images and a variety of object categories. Detection results confirm the effectiveness and robustness of the learned parts. 1.

Human cognition is unique in the way in which it relies on combinatorial (or compositional) structures. Language provides ample evidence for the existence of combinatorial structures, but they can also be found in visual cognition. To understand the neural basis of human cognition, it is therefore essential to understand how combinatorial structures can be instantiated in neural terms. In his recent book on the foundations of language, Jackendoff formulated four fundamental problems for a neural instantiation of combinatorial structures: the massiveness of the binding problem, the problem of 2, the problem of variables and the transformation of combinatorial structures from working memory to long-term memory. This paper aims to show that these problems can be solved by means of neural ‘blackboard ’ architectures. For this purpose, a neural blackboard architecture for sentence structure is presented. In this architecture, neural structures that encode for words are temporarily bound in a manner that preserves the structure of the sentence. It is shown that the architecture solves the four problems presented by Jackendoff. The ability of the architecture to instantiate sentence structures is illustrated with examples of sentence complexity observed in human language performance. Similarities exist between the architecture for sentence structure and blackboard architectures for combinatorial structures in visual cognition, derived from the structure of the visual cortex. These architectures are briefly discussed, together with an example of a combinatorial structure in which the blackboard architectures for language and vision are combined. In this way, the architecture for language is grounded in perception. 2 Content

Are minds really dynamical or are they really symbolic? Because minds are bundles of computations, and because computation is always a matter of interpretation of one system by another, minds are necessarily symbolic. Because minds, along with everything else in the universe, are physical, and insofar as the laws of physics are dynamical, minds are necessarily dynamical systems. Thus, the short answer to the opening question is “yes.” It makes sense to ask further whether some of the computations that constitute a human mind are constrained by functional, algorithmic, or implementational factors to be essentially of the discrete symbolic variety (even if they supervene on an apparently continuous dynamical substrate). I suggest that here too the answer is “yes” and discuss the need for such discrete, symbolic cognitive computations in communication-related tasks.

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 human capacity for visual object recognition is characterized by a number of properties that are jointly very challenging to explain. Among the most notable is that the visual representation of shape is invariant with (i.e., insensitive to) the location of the image in the visual field (Biederman & Cooper, 1991a), the size of the image (Biederman & Cooper, 1992), left–right (i.e., mirror) reflection (Biederman & Cooper, 1991a; Davidoff & Warrington, 2001), and some rotations in depth (Biederman & Gerhardstein, 1993, 1995; but see Tarr & Bülthoff, 1995). Object recognition is also remarkably robust to variations in shape (Davidoff & Warrington, 1999). For example, a child’s drawing of a car may be easily recognizable as a car, even if it resembles neither any particular real car nor any road-worthy object. At the same time, object recognition is sensitive to rotations about the line of sight

Abstract. To learn a visual code in an unsupervised manner, one may attempt to capture those features of the stimulus set that would contribute significantly to a statistically efficient representation (as dictated, e.g., by the Minimum Description Length principle). Paradoxically, all the candidate features in this approach need to be known before statistics over them can be computed. This paradox may be circumvented by confining the repertoire of candidate features to actual scene fragments, which resemble the “what+where ” receptive fields found in the ventral visual stream in primates. We describe a single-layer network that learns such fragments from unsegmented raw images of structured objects. The learning method combines fast imprinting in the feedforward stream with lateral interactions to achieve single-epoch unsupervised acquisition of spatially localized features that can support systematic treatment of structured objects [1]. 1 A paradox and some ways of resolving it

A fundamental aspect of human intelligence is the ability to represent and reason about relations. Examples of relational thinking include our ability to appreciate analogies between different objects or events (Gentner, 1983; Holyoak & Thagard, 1995), our ability to apply abstract rules in novel situations (e.g., Smith, Langston & Nisbett, 1992), our ability to understand and learn language (e.g., Kim, Pinker, Prince & Prasada, 1991), our

d processing (Solan et al., 2003). The new computational model gives up the logicism of generative grammar in favor of information-theoretic learning of distributed construction patterns. The structure and the meaning of a sentence (which can be thought of as the proverbial elephant groped by the blind men) are thus represented by the chorus of responses of construction detectors, which can be further processed by methods that are being worked out for another cognitive domain with similar computational needs: vision (Edelman and Intrator, 2003). Recent empirical findings indicate that (1) langue, and not merely parole, is imperfect (Chipere, 2001), just as the other faculties of the mind are, (2) people can handle only very shallow true recursion or center embedding (MacDonald and Christiansen, 2002), and (3) language is more formulaic than creative (Wray, 2002). The newly emerging computational work shows also that (4) linguistic knowledge can be learned from scratch, and (5) relianc

Summary. This paper presents our recently developed approach to constructing a hierarchical representation of visual input that aims to enable recognition and detection of a large number of object categories. Inspired by the principles of efficient indexing, robust matching, and ideas of compositionality, our approach learns a hierarchy of spatially flexible compositions, i.e. parts, in an unsupervised, statistics-driven manner. Starting with simple, frequent features, we learn the statistically most significant compositions (parts composed of parts), which consequently define the next layer. Parts are learned sequentially, layer after layer, optimally adjusting to the visual data. Lower layers are learned in a category-independent way to obtain complex, yet sharable visual building blocks, which is a crucial step towards a scalable representation. Higher layers of the hierarchy, on the other hand, are constructed by using specific categories, achieving a category representation with a small number of highly generalizable parts that gained their structural flexibility through composition within the hierarchy. Built in this way, new categories can be efficiently and continuously added to the system by adding a small number of parts only in the higher layers. The approach is demonstrated on a large collection of images and a variety of object categories. 1