Plaut Lab - Carnegie Mellon University


Normal and Impaired Word Reading

Much of my early work focused on word reading, both in normal skilled readers and in brain-damaged patients with acquired reading disorders. Word reading is a particularly informative domain for studying cognitive processes because it involves learning to relate multiple sources of information—visual (orthographic), phonological, and semantic—in a highly skilled manner. My colleagues and I have developed artificial neural-network (connectionist) models that exhibit many of the central characteristics of skilled reading, including the influences of word frequency and spelling-sound consistency on the time to pronounce words and the ability to pronounce word-like nonsense letter strings (e.g., MAVE) and to distinguish them from real words in lexical decision tasks (Plaut, McClelland, Seidenberg & Patterson, 1996). When the models are damaged in various ways, they exhibit the major forms of acquired dyslexia, including deep dyslexia, in which patients make semantic errors in reading aloud (e.g., misreading YACHT as “boat”; Plaut & Shallice, 1993) and surface dyslexia, in which patients produce regularization errors to exception words (e.g., misreading YACHT as “yatched”; Woollams, Lambon Ralph, Plaut & Patterson, 2007). Moreover, retraining the damaged models yields patterns of recovery and generalization that are qualitatively similar to those found in cognitive rehabilitation studies and has, in one instance (Plaut, 1996), generated a specific prediction concerning the design of more effective therapy for patients that later received direct empirical support (Kiran & Thompson, 2003, JSLHR).

Derivational and Inflectional Morphology

Traditional theories posit that complex words are composed of discrete units called morphemes that contribute systematically to their meanings (e.g., TEACH+ER, GOVERN+MENT), but some words are awkward on this account (e.g., DRESS+ER is not someone who dresses; MOTH+ER, FATH+ER, SIST+ER, BROTH+ER and all agents but the remaining parts are not coherent units). On a distributed connectionist approach, however, morphology reflects a learned sensitivity to the graded degree of systematicity among the surface forms of words and their meanings, without the need to posit discrete segmentation. Explicit simulations demonstrate that, in accordance with empirical findings (e.g., Velan, Frost, Deustch & Plaut, 2005), the degree of sensitivity to apparent morphological structure in the absence of semantic similarity (e.g., BROTH+ER) depends on the overall morphological richness of the language as a whole (Plaut & Gonnerman, 2000). More generally, insights drawn from the connectionist perspective on morphology and its debate with “rule-based” accounts---in particular, the English past-tense system---have been assimilated into many areas in the study of language, changing the focus of research from abstract characterizations of linguistic competence to an emphasis on the role of the statistical structure of language in acquisition and processing (Seidenberg & Plaut, 2014).

Semantics and Word Comprehension

A longstanding debate regarding the representation of semantic knowledge is whether such knowledge is represented in a single, amodal system or whether it is organized into multiple subsystems based on modality of input or type of information. A distributed connectionist perspective offers a middle ground, in which semantic representations develop under the pressure of learning to mediate between multiple input and output modalities in performing various tasks, under a constraint to minimize connection length (and, hence, overall axon volume). An implemented model provides a quantitative account of optic aphasia---a selective deficit in naming visually presented objects---following damage to connections from vision to regions of semantics near phonology (Plaut, 2002). Additional implementations of the process by which visual representations activate semantics account for 1) detailed patterns of semantic priming and how these vary across individuals over the course of development (Plaut & Booth, 2000); 2) distinct patterns of impairment in word and picture comprehension reflecting “access” versus “degraded-store” deficits (Gotts & Plaut, 2002); and 3) the influence of orthographic and semantic variables on the shape and magnitude of the N400 evoked response potential (ERP), which is thought to reflect “semantic integration” and corresponds simply to overall semantic activation in the model (Laszlo & Plaut, 2012).

Sequential Behavior

In everyday tasks, selecting actions in the proper sequence requires a continuously updated representation of temporal context. Many existing models address this problem by positing a hierarchy of processing units, mirroring the roughly hierarchical structure of naturalistic tasks themselves. Although intuitive, such an approach has led to a number of difficulties, including a reliance on overly rigid sequencing mechanisms and a limited ability to address both learning and context sensitivity in behavior. A sequential neural network, by contrast, can to deal flexibly with a complex set of sequencing constraints, encoding contextual information at multiple time-scales within a single, distributed internal representation (Botvinick & Plaut, 2004). The model not only accounts for skilled action performance, but also everyday “slips of action” that normal individuals commit under distraction, as well as more severe degradation in performance following damage, as observed in ideational apraxia. An analogous model in the domain of language acquisition and processing. accounts for the integration of semantic and syntactic constraints on sentence processing (Rohde & Plaut, 1999). Finally, the same type of model, at a shorter timescale, provides a parsimonious account for numerous benchmark phenomena in the domain of immediate serial recall (Botvinick & Plaut, 2006), including data that have been considered to preclude the application of neural networks in this domain. Unlike most competing accounts, the model deals naturally with findings concerning the role of background knowledge in serial recall, and makes contact with relevant neuroscientific data.

Neural Representation of Faces and Words

The neural mechanisms supporting visual recognition of faces, words, and other objects are increasingly conceptualized as a distributed but integrated system that become organized gradually over the course of development, rather than as a set of individual, specialized regions subserving particular visual domains (Behrmann & Plaut, 2013). In understanding the emergence of this organization, we adopt a specific theoretical perspective in which visual recognition involves topographically-constrained cooperation and competition among multiple, interacting regions, each of which is only partially selective for a specific domain. When applied to faces and words in an explicit computational simulation (Plaut & Behrmann, 2011), these domains compete to be near high-acuity visual information in each hemisphere; words become more left-lateralized to cooperate with language-related information and, in response, faces subsequently become more right-lateralized. The account thus makes specific and otherwise unexpected predictions—supported by subsequent empirical studies (e.g., Behrmann & Plaut, 2014; Nestor, Behrmann & Plaut, 2013; Nestor, Plaut & Behrmann, 2013)—concerning the co-mingling of these two seemingly unrelated domains over the course of development, in neurophysiological measures of recognition in both children and adults, and in graded patterns of impairment in both domains following unilateral brain damage. The research offers a novel theoretical perspective that has broad implications for theories of normal and atypical cognitive and neural development, and for instruction and remediation.

New Projects

Semantic ambiguity. The meanings of most words depend on the context in which they occur (e.g., vs. BANK). Developing a theory of how comprehension of semantically ambiguous words are understood is a critical aspect of any theory of word or discourse comprehension. However, success to date has been limited by discrepancies in the effects of relatedness of meaning observed within and between tasks. Further, existing accounts are underspecified, narrow in scope, and mutually inconsistent. The current work introduces the semantic settling dynamics account of semantic ambiguity resolution, in which the discrepant effects are explained by the temporal settling dynamics in semantics within a neural network, and how these dynamics interact with the semantic representations of ambiguous words over time. This account stands as an alternative to one based on the configuration of the decision system across tasks (Hino, Pexman, & Lupker, 2006, Journal of Memory and Language). The proposed account reconciles a wide body of disparate results within a single unified mechanistic account, is supported by initial investigations that vary processing time to modulate semantic ambiguity effects, and generates targeted predictions for future computational, neural, and behavioral research. Statistical learning. Statistical learning is often cast as a means of discovering the units of perception, such as words and objects, and representing them as explicit "chunks". However, entities are not undifferentiated wholes but often contain parts that contribute systematically to their meanings. Studies of incidental auditory or visual statistical learning suggest that, as participants learn about wholes they become insensitive to parts embedded within them (Fiser & Aslin, 2005; Giroux & Rey, 2009), but this seems difficult to reconcile with a broad range of findings in which parts and wholes work together to contribute to behavior. In the current work, we adopt a computational approach, based on learning in artificial neural networks, that is capable of capturing statistical structure at multiple levels of representation simultaneously and yet eschews the notion of explicit chunks. Rather, the extent to which a particular subset of the input in a particular context is represented in a coherent manner is a matter of degree, and the extent to which structure at one level of analysis cooperates or competes with structure at other levels is not prespecified but arises naturally through incidental learning. We show that the approach accounts for a wide range of findings concerning the relationship between parts and wholes in auditory and visual statistical learning, including some previously thought to be problematic for neural network approaches. N400. The study of the N400 event-related brain potential has provided fundamental insights into the nature of real-time comprehension processes, and its amplitude is modulated by a wide variety of stimulus and context factors. It is generally thought to reflect the difficulty of semantic access, but formulating a precise characterization of this process has proved difficult. Laszlo and colleagues (Laszlo & Plaut, 2012, Brain and Language, 120, 271-281; Laszlo & Armstrong, 2014, Brain and Language, 132, 22-27) used physiologically constrained neural networks to model the N400 as transient over-activation within semantic representations, arising as a consequence of the distribution of excitation and inhibition within and between cortical areas. The current work extends this approach to successfully model effects on both N400 amplitudes and behavior of word frequency, semantic richness, repetition, semantic and associative priming, and orthographic neighborhood size. The account is argued to be preferable to one based on "semantic prediction error" (Rabovsky & McRae, 2014, Cognition, 132, 68-98) for a number of reasons, the most fundamental of which is that the current model actually produces N400-like waveforms in its real-time activation dynamics. Dorsal object representations. The cortical visual system is almost universally thought to be segregated into two anatomically and functionally distinct pathways: a ventral occipito-temporal pathway that subserves object perception, and a dorsal occipito-parietal pathway that subserves object localization and visually guided action. Accumulating evidence from both human and non-human primate studies, however, challenges this binary distinction and suggests that regions in the dorsal pathway contain object representations that are independent of those in ventral cortex and that play a functional role in object perception. We are exploring the nature of dorsal object representations through a combination of behavioral, neuropsychological, neuroimaging, and computational work. We propose a graded functional account of the anatomical organization, functional contributions and origins of these representations in the service of perception and action.
  • Freud, E., Plaut, D.C., and Behrmann, M. (2016). "What" is happening in the dorsal visual pathway. Trends in Cognitive Sciences, 20, 773-784. doi:10.1016/j.tics.2016.08.003
  • Freud, E., Culham, J., Plaut, D.C. and Behrmann, M. (submitted). The large-scale organization of shape processing in the ventral and dorsal pathways. eLife.