Shimon Edelman

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. 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]

Lasnik’s review of the Minimalist program in syntax [1] offers cognitive scientists help in navigating some of the arcana of the current theoretical thinking in transformational generative grammar. One may observe, however, that this journey is more like a taxi ride gone bad than a free tour: it is the driver who decides on the itinerary, and questioning his choice may get you kicked out. Meanwhile, the meter in the cab of the generative theory of grammar is running, and has been since the publication of Chomsky’s Syntactic Structures in 1957. The fare that it ran up is none the less daunting for the detours made in his Aspects of Theory of Syntax in 1965, Government and Binding in 1981, and now The Minimalist Program, in 1995. Paraphrasing Winston Churchill, it seems like never in the field of cognitive science was so much owed by so many of us to so few. For most of us in the cognitive sciences this situation will appear quite benign, if we realize that it is the generative linguists who should by rights be paying this bill. The reason for that is simple and is well known in the philosophy of science: putting forward a theory is like taking out a loan, to be repayed by gleaning an empirical basis for it; theories that fail to do so are declared bankrupt. In the sciences of the mind, this maxim translates into the need to demonstrate the psychological, and, eventually, the neurobiological, reality of the theoretical constructs. Many examples of this process can be found in the study of human vision, where, as in language, direct observation of the underlying mechanisms is difficult; for instance, the concept of multiple parallel spatial frequency channels, introduced in the late 1960s, was completely vindicated by purely behavioral means over the following decade; see, e.g., [2]. In linguistics, the nature of the requisite evidence is well described by Townsend and Bever: “What do we test today if we want to explore the behavioral implications of syntax?.

A metaphor that has dominated linguistics for the entire duration of its existence as a discipline views sentences as edifices consisting of Lego-like building blocks. It is assumed that each sentence is constructed (and, on the receiving end, parsed) ab novo, starting (ending) with atomic constituents, to logical semantic specifications, in a recursive process governed by a few precise algebraic rules. The assumptions underlying the Lego metaphor, as it is expressed in generative grammar theories, are: (1) perfect regularity of what Saussure called langue, (2) infinite potential recursivity of syntactic structures, (3) unlimited human capacity for linguistic creativity, (4) the impossibility of acquiring structural knowledge from examples, and (5) the impossibility of such knowledge being stored in a memory-intensive form (ensembles of exemplars).

Two experiments investigated whether infants represent goal‐directed actions of others in a way that allows them to draw inferences to unobserved states of affairs (such as unseen goal states or occluded obstacles). We measured looking times to assess violation of infants' expectations upon perceiving either a change in the actions of computer‐animated figures or in the context of such actions. The first experiment tested whether infants would attribute a goal to an action that they had not seen completed. The second experiment tested whether infants would infer from an observed action the presence of an occluded object that functions as an obstacle. The looking time patterns of 12‐month‐olds indicated that they were able to make both types of inferences, while 9‐month‐olds failed in both tasks. These results demonstrate that, by the end of the first year of life, infants use the principle of rational action not only for the interpretation and prediction of goal‐directed actions, but also for making productive inferences about unseen aspects of their context. We discuss the underlying mechanisms that may be involved in the developmental change from 9 to 12 months of age in the ability to infer hypothetical (unseen) states of affairs in teleological action representations.

Supposing the symbol system postulated by Barsalou is perceptual through and through -- what then? The target article outlines an intriguing and exciting theory of cognition in which (1) wellspecified, event- or object-linked percepts assume the role traditionally allotted to abstract and arbitrary symbols, and (2) perceptual simulation is substituted for processes traditionally believed to require symbol manipulation, such as deductive reasoning. We take a more extreme stance on the role of perception (in particular, vision) in shaping cognition, and propose, in addition to Barsalou's postulates, that (3) spatial frames, endowed with a perceptual structure not unlike that of the retinotopic space, pervade all sensory modalities and are used to support compositionality.

The problem of representing the spatial structure of images, which arises in visual object processing, is commonly described using terminology borrowed from propositional theories of cognition, notably, the concept of compositionality. The classical propositional stance mandates representations composed of symbols, which stand for atomic or composite entities and enter into arbitrarily nested relationships. We argue that the main desiderata of a representational system—productivity and systematicity—can (indeed, for a number of reasons, should) be achieved without recourse to the classical, proposition‐like compositionality. We show how this can be done, by describing a systematic and productive model of the representation of visual structure, which relies on static rather than dynamic binding and uses coarsely coded rather than atomic shape primitives.

An image of a face depends not only on its shape, but also on the viewpoint, illumination conditions, and facial expression. A face recognition system must overcome the changes in face appearance induced by these factors. This paper investigate two related questions: the capacity of the human visual system to generalize the recognition of faces to novel images, and the level at which this generalization occurs. We approach this problems by comparing the identi cation and generalization capacity for upright and inverted faces. For upright faces, we found remarkably good generalization to novel conditions. For inverted faces, the generalization to novel views was signi cantly worse for both new illumination and viewpoint, although the performance on the training images was similar to the upright condition. Our results indicate that at least some of the processes that support generalization across viewpoint and illumination are neither universal (because subjects did not generalize as easily for inverted faces as for upright ones), nor strictly objectspeci c (because in upright faces nearly perfect generalization was possible from a single view, by itself insu cient for building a complete object-speci c model). We propose that generalization in face recognition occurs at an intermediate level that is applicable to a class of objects, and that at this level upright and inverted faces initially constitute distinct object classes.

Visual objects can be represented by their similarities to a small number of reference shapes or prototypes. This method yields low-dimensional (and therefore computationally tractable) representations, which support both the recognition of familiar shapes and the categorization of novel ones. In this note, we show how such representations can be used in a variety of tasks involving novel objects: viewpoint-invariant recognition, recovery of a canonical view, estimation of pose, and prediction of an arbitrary view. The unifying principle in all these cases is the representation of the view space of the novel object as an interpolation of the view spaces of the reference shapes.

Intelligent systems are faced with the problem of securing a principled (ideally, veridical) relationship between the world and its internal representation. I propose a unified approach to visual representation, addressing both the needs of superordinate and basic-level categorization and of identification of specific instances of familiar categories. According to the proposed theory, a shape is represented by its similarity to a number of reference shapes, measured in a high-dimensional space of elementary features. This amounts to embedding the stimulus in a low-dimensional proximal shape space. That space turns out to support representation of distal shape similarities which is veridical in the sense of Shepard's (1968) notion of second-order isomorphism (i.e., correspondence between distal and proximal similarities among shapes, rather than between distal shapes and their proximal representations). Furthermore, a general expression for similarity between two stimuli, based on comparisons to reference shapes, can be used to derive models of perceived similarity ranging from continuous, symmetric, and hierarchical, as in the multidimensional scaling models (Shepard, 1980), to discrete and non-hierarchical, as in the general contrast models (Tversky, 1977; Shepard and Arabie, 1979)

We tested the hypothesis that more frequent exposure to multiword phrases results in deeper entrenchment of their representations, by examining the performance of subjects of different religiosity in the recognition of briefly presented liturgical and secular phrases drawn from several frequency classes. Three of the sources were prayer texts that religious Jews are required to recite on a daily, weekly, and annual basis, respectively; two others were common and rare expressions encountered in the general secular Israeli culture. As expected, linear dependence of recognition score on frequency was found for the religious subjects (being most pronounced for men, who are usually more observant than women); both religious and secular subjects performed better on common than on rare general culture items. Our results support the notion of graded entrenchment introduced by Langacker and shared by several cognitive linguistic theories of language comprehension and production

We introduce a set of biologically and computationally motivated design choices for modeling the learning of language, or of other types of sequential, hierarchically structured experience and behavior, and describe an implemented system that conforms to these choices and is capable of unsupervised learning from raw natural-language corpora. Given a stream of linguistic input, our model incrementally learns a grammar that captures its statistical patterns, which can then be used to parse or generate new data. The grammar constructed in this manner takes the form of a directed weighted graph, whose nodes are recursively (hierarchically) defined patterns over the elements of the input stream. We evaluated the model in seventeen experiments, grouped into five studies, which examined, respectively, (a) the generative ability of grammar learned from a corpus of natural language, (b) the characteristics of the learned representation, (c) sequence segmentation and chunking, (d) artificial grammar learning, and (e) certain types of structure dependence. The model's performance largely vindicates our design choices, suggesting that progress in modeling language acquisition can be made on a broad front—ranging from issues of generativity to the replication of human experimental findings—by bringing biological and computational considerations, as well as lessons from prior efforts, to bear on the modeling approach.

Language is a rewarding field if you are in the prediction business. A reader who is fluent in English and who knows how academic papers are typically structured will readily come up with several possible guesses as to where the title of this section could have gone, had it not been cut short by the ellipsis. Indeed, in the more natural setting of spoken language, anticipatory processing is a must: performance of machine systems for speech interpretation depends critically on the availability of a good predictive model of how utterances unfold in time (Baker, 1975; Jelinek, 1990; Goodman, 2001), and there is strong evidence that prospective uncertainty affects human sentence processing too (Jurafsky, 2003; Hale, 2006; Levy, 2008). The human ability to predict where the current utterance is likely to be going is just another adaptation to the general pressure to anticipate the future (Hume, 1748; Dewey, 1910; Craik, 1943), be it in perception, thinking, or action, which is exerted on all cognitive systems by evolution (Dennett, 2003). Look-ahead in language is, however, special in one key respect: language is a medium for communication, and in communication the most interesting (that is, informative) parts of the utterance that the speaker is working through are those that cannot be predicted by the listener ahead of time.

Unsupervised statistical learning is the standard setting for the development of the only advanced visual system that is both highly sophisticated and versatile, and extensively studied: that of monkeys and humans. In this extended abstract, we invoke philosophical observations, computational arguments, behavioral data and neurobiological findings to explain why computer vision researchers should care about (1) unsupervised learning, (2) statistical inference, and (3) the visual brain. We then outline a neuromorphic approach to structural primitive learning motivated by these considerations, survey a range of neurobiological findings and behavioral data consistent with it, and conclude by mentioning some of the more challenging directions for future research.

By what empirical means can a person determine whether he or she is presently awake or dreaming? Any conceivable test addressing this question, which is a special case of the classical metaphysical doubting of reality, must be statistical (for the same reason that empirical science is, as noted by Hume). Subjecting the experienced reality to any kind of statistical test (for instance, a test for bizarreness) requires, however, that a set of baseline measurements be available. In a dream, or in a simulation, any such baseline data would be vulnerable to tampering by the same processes that give rise to the experienced reality, making the outcome of a reality test impossible to trust. Moreover, standard cryptographic defenses against such tampering cannot be relied upon, because of the potentially unlimited reach of reality modification within a dream, which may range from the integrity of the verification keys to the declared outcome of the entire process. In the face of this double predicament, the rational course of action is to take reality at face value. The predicament also has some intriguing corollaries. In particular, even the most revealing insight that a person may gain into the ultimate nature of reality (for instance, by attaining enlightenment in the Buddhist sense) is ultimately unreliable, for the reasons just mentioned. At the same time, to adhere to this principle, one has to be aware of it, which may not be possible in various states of reduced or altered cognitive function such as dreaming or religious experience. Thus, a subjectively enlightened person may still lack the one truly important piece of the puzzle concerning his or her existence.