Shimon Edelman

The computational program for theoretical neuroscience initiated by Marr and Poggio (1977) calls for a study of biological information processing on several distinct levels of abstraction. At each of these levels — computational (defining the problems and considering possible solutions), algorithmic (specifying the sequence of operations leading to a solution) and implementational — significant progress has been made in the understanding of cognition. In the past three decades, computational principles have been discovered that are common to a wide range of functions in perception (vision, hearing, olfaction) and action (motor control). More recently, these principles have been applied to the analysis of cognitive tasks that require dealing with structured information, such as visual scene understanding and analogical reasoning. Insofar as language relies on cognition-general principles and mechanisms, it should be possible to capitalize on the recent advances in the computational study of cognition by extending its methods to linguistics.

Nearest-neighbor correlation-based similarity computation in the space of outputs of complex-type receptive elds can support robust recognition of 3D objects. Our experiments with four collections of objects resulted in mean recognition rates between 84% and 94%, over a 40 40 range of viewpoints, centered on a stored canonical view and related to it by rotations in depth. This result has interesting implications for the design of a front end to an arti cial object recognition system, and for the understanding of the faculty of object recognition in primate vision

We describe a method for automatic word sense disambiguation using a text corpus and a machine- readable dictionary (MRD). The method is based on word similarity and context similarity measures. Words are considered similar if they appear in similar contexts; contexts are similar if they contain similar words. The circularity of this definition is resolved by an iterative, converging process, in which the system learns from the corpus a set of typical usages for each of the senses of the polysemous word listed in the MRD. A new instance of a polysemous word is assigned the sense associated with the typical usage most similar to its context. Experiments show that this method performs well, and can learn even from very sparse training data.

The distributional principle according to which morphemes that occur in identical contexts belong, in some sense, to the same category [1] has been advanced as a means for extracting syntactic structures from corpus data. We extend this principle by applying it recursively, and by using mutual information for estimating category coherence. The resulting model learns, in an unsupervised fashion, highly structured, distributed representations of syntactic knowledge from corpora. It also exhibits promising behavior in tasks usually thought to require representations anchored in a grammar, such as systematicity.

It is proposed to conceive of representation as an emergent phenomenon that is supervenient on patterns of activity of coarsely tuned and highly redundant feature detectors. The computational underpinnings of the outlined concept of representation are (1) the properties of collections of overlapping graded receptive fields, as in the biological perceptual systems that exhibit hyperacuity-level performance, and (2) the sufficiency of a set of proximal distances between stimulus representations for the recovery of the corresponding distal contrasts between stimuli, as in multidimensional scaling. The present preliminary study appears to indicate that this concept of representation is computationally viable, and is compatible with psychological and neurobiological data

differentiaily rated pairwise similarity when confronted with two pairs of objects, each revolving in a separate window on a computer screen. Subject data were pooled using individually weighted MDS (ref. 11; in all the experiments, the solutions were consistent among subjects). In each trial, the subject had to select among two pairs of shapes the one consisting of the most similar shapes. The subjects were allowed to respond at will; most responded within 10 sec. Proximity (that is, perceived similarity) tables derived from the judgments were processed to verify their degree of transitivity (4% of all triplets were found intransitive) and then submitted to MDS. In the long-term memory (LTM) variant of this experiment, the subjects were first trained to associate a label (a three-letter nonsensical string, such as "BON" or "POM") with each object and then carried out the pairs of pairs comparison task from memory, prompted by the object labels rather than by the objects themselves. Six subjects participated in each of the two LTM experiments (Star and Triangle). The subjects were taught each shape in a separate session and had to discriminate between that shape and six similar nontargets from various viewpoints. Training continued until the recognition rate reached 90%, over a period of several days. The subjects were never exposed to more than one target in one session and were not told the ultimate purpose of the experiment. After 2 to 3 days of rest, they were tested with questions such as: "is the BON more similar to POM than TOC to ROX?", for all pairs of pairs of stimuli. In the LTM experiments, 8% of the..

Construction-based approaches to syntax (Croft, 2001; Goldberg, 2003) posit a lexicon populated by units of various sizes, as envisaged by (Langacker, 1987). Constructions may be specified completely, as in the case of simple morphemes or idioms such as take it to the bank, or partially, as in the expression what’s X doing Y?, where X and Y are slots that admit fillers of particular types (Kay and Fillmore, 1999). Constructions offer an intriguing alternative to traditional rule-based syntax by hinting at the extent to which the complexity of language can stem from a rich repertoire of stored, more or less entrenched (Harris, 1998) representations that address both syntactic and semantic issues, and encompass, in addition to general rules, “totally idiosyncratic forms and patterns of all intermediate degrees of generality” (Langacker, 1987, p.46). Because constructions are by their very nature language-specific, the question of acquisition in Construction Grammar is especially poignant. We address this issue by offering an unsupervised algorithm that learns constructions from raw corpora.

We examined the role of fitness, commonly assumed without proof to be conferred by the mastery of language, in shaping the dynamics of language evolution. To that end, we introduced island migration (a concept borrowed from population genetics) into the shared lexicon model of communication (Nowak et al., 1999). The effect of fitness linear in language coherence was compared to a control condition of neutral drift. We found that in the neutral condition (no coherence-dependent fitness) even a small migration rate – less than 1% – suffices for one language to become dominant, albeit after a long time. In comparison, when fitness-based selection is introduced, the subpopulations stabilize quite rapidly to form several distinct languages. Our findings support the notion that language confers increased fitness. The possibility that a shared language evolved as a result of neutral drift appears less likely, unless migration rates over evolutionary times were extremely small.

Converging evidence from anatomical studies (Maunsell, 1983) and functional analyses (Hubel & Wisesel, 1968) of the nervous system suggests that the feed-forward pathway of the mammalian perceptual system follows a largely hierarchic organization scheme. This may be because hierarchic structures are intrinsically more viable and thus more likely to evolve (Simon, 2002). But it may also be because objects in our environment have a hierarchic structure and the perceptual system has evolved to match it. We conducted a behavioral experiment to investigate the effect of the degree of hierarchy of the generative probabilistic structure in categorization. We generated one set of stimuli using a hierarchic underlying probability distribution, and another set according to a non-hierarchic one. Participants were instructed to categorize these images into one of the two possible categories a. Our results suggest that participants perform more accurately in the case of hierarchically structured stimuli.

We compare our model of unsupervised learning of linguistic structures, ADIOS [1, 2, 3], to some recent work in computational linguistics and in grammar theory. Our approach resembles the Construction Grammar in its general philosophy (e.g., in its reliance on structural generalizations rather than on syntax projected by the lexicon, as in the current generative theories), and the Tree Adjoining Grammar in its computational characteristics (e.g., in its apparent affinity with Mildly Context Sensitive Languages). The representations learned by our algorithm are truly emergent from the (unannotated) corpus data, whereas those found in published works on cognitive and construction grammars and on TAGs are hand-tailored. Thus, our results complement and extend both the computational and the more linguistically oriented research into language acquisition. We conclude by suggesting how empirical and formal study of language can be best integrated.

Variation set structure — partial alignment of successive utterances in child-directed speech — has been shown to correlate with progress in the acquisition of syntax by children. The present study demonstrates that arranging a certain proportion of utterances in a training corpus in variation sets facilitates word segmentation and phrase structure learning in miniature artifi- cial languages by adults. Our findings have implications for understanding the mechanisms of L1 acquisition by children, and for the development of more efficient algorithms for automatic language acquisition, as well as better methods for L2 instruction.

A theory of representation is incomplete if it states “representations areX” whereXcan be symbols, cell assemblies, functional states, or the flock of birds fromTheaetetus, without explaining the nature of the link between the universe ofXs and the world. Amit's thesis, equating representations with reverberations in Hebbian cell assemblies, will only be considered a solution to the problem of representation when it is complemented by a theory of how a reverberation in the brain can be a representation of anything.

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]

We report a quantitative analysis of the cross-utterance coordination observed in child-directed language, where successive utterances often overlap in a manner that makes their constituent structure more prominent, and describe the application of a recently published unsupervised algorithm for grammar induction to the largest available corpus of such language, producing a grammar capable of accepting and generating novel wellformed sentences. We also introduce a new corpus-based method for assessing the precision and recall of an automatically acquired generative grammar without recourse to human judgment. The present work sets the stage for the eventual development of more powerful unsupervised algorithms for language acquisition, which would make use of the coordination structures present in natural child-directed speech.

The learnability of features and their dependence on task and context do not rule out the possibility that primitives used for constructing new features are as small as pixels, nor that they are as large as object parts, or even entire objects. In fact, the simplest approach to feature acquisition may be to treat objects not as if they are composed of unknown primitives according to unknown rules, but rather as if they are what they seem: patterns of atomic features, standing in various similarity relationships to other objects, which serve as holistic features.

Beer ’s paper devotes much energy to buttressing the walls of Castle Dynamic and dredging its moat in the face of what some of its dwellers perceive as a besieging army chanting “no cognition without representation”. The divide is real, as attested by the contrast between titles such as “Intelligence without representation” and “In defense of representation”, to pick just one example from each side. It is, however, not too late for people from both sides of the moat to meet on the drawbridge and see if all that energy can be put to a better use. The parley can be organized around an attempt to identify those attributes of representations that both sides may find useful. In my view, one such attribute is the capacity for hierarchical abstraction, an aspect of the representational approach not actually mentioned by Beer. The capacity for abstraction is gained automatically by those theoretical frameworks that opt for the explanatory benefits of representation in understanding cognition. Moreover, without appropriately structured mediating states, a cognitive system would be incapable of dealing with complex reality. Thus, hierarchical abstraction is useful not just for the cognitive scientists in their attempts to scale up their understanding of cognition: it is also indispensable for cognitive systems that aspire to scale up their understanding of the world

We compare our model of unsupervised learning of linguistic structures, ADIOS [1], to some recent work in computational linguistics and in grammar theory. Our approach resembles the Construction Grammar in its general philosophy (e.g., in its reliance on structural generalizations rather than on syntax projected by the lexicon, as in the current generative theories), and the Tree Adjoining Grammar in its computational characteristics (e.g., in its apparent affinity with Mildly Context Sensitive Languages). The representations learned by our algorithm are truly emergent from the (unannotated) corpus data, whereas those found in published works on cognitive and construction grammars and on TAGs are hand-tailored. Thus, our results complement and extend both the computational and the more linguistically oriented research into language acquisition. We conclude by suggesting how empirical and formal study of language can be best integrated.

Representational systems need to use symbols as internal stand-ins for distal quantities and events. Barsalou's ideas go a long way towards making the symbol system theory of representation more appealing, by delegating one critical part of the representational burden to image-like entities. The target article, however, leaves the other critical component of any symbol system theory underspecified. We point out that the binding problem can be alleviated if a perceptual symbol system is made to rely on image-like entities not only for grounding the constituent symbols, but also for composing these into structures