Commentary on Thomas
Abstract: 60 words
Main Text: 1386 words
References: 377 words
Total Text: 1867 words
Thomas and Karmiloff-Smith (2002) (hereafter, TKS) have done the field
an important service by calling attention to the fact that lesions to
a developing brain can have different consequences than lesions to
adult brains. They are entirely correct in pointing out that adaptive
learning processes will shape, not only the acquired functions of
those parts of the brain themselves directly affected by the damage,
but also other parts of the brain. In general, the article makes a
point we heartily agree with, namely that the standard logic of
neuropsychological interpretation should not be applied uncritically
to the interpretation of developmental disorders.
In our view (see also Plaut, 1995), the standard logic of neuropsychological interpretation cannot be applied uncritically to the interpretation of any disorders, whether or not they are developmental disorders. By the phrase 'the standard logic of neuropsychological interpretation' we mean the reliance on a double dissociation between performance on materials from two different experimenter-defined categories to infer that normal performance relies on separate modules specialized for processing the different categories of materials. This logic has repeatedly been used in both the adult neuropsychology literature and in the literature on developmental disorders discussed by TKS (for discussions, see Shallice, 1988 and Plaut, 1995). For example, a double-dissociation between living things and artifacts in picture naming and property verification has been used to argue for separate modules for different semantic categories (Warrington and McCarthy, 1987), and a double-dissociation in reading abstract vs. concrete words been used to argue for separate modules for abstract vs concrete nouns (Warrington, 1981). Closer to the example used by TKS, a double dissociation in production of the past tenses of exception words vs. regular words has been used to argue for separate brain mechanisms for words and rules (Pinker, 1991; Pinker and Ullman, 2002).
In all these cases, connectionist/parallel-distributed processing models (Rumelhart, McClelland, and the PDP Research Group, 1986) have provided alternatives to the standard interpretations of these double dissociations (Farah and McClelland, 1991; Plaut, 1995; Joanisse and Seidenberg, 1999). In general, these models take the following form: A single integrated and interactive system is used for processing items of both categories. Due to item characteristics that co-vary with category membership, performance on items from one category depends more on one part of the system, while performance on items from the other category depends more on another part of the system. For example, Plaut (1995) suggested that concrete and abstract words may differ in the number of semantic features. He trained a network with feed-forward and recurrent connections to map both concrete and abstract words from orthography to semantics. The concrete words contained more semantic features, so they used the recurrent connections more effectively and were less dependent on the feed-forward connections. Lesioning the feed-forward connections produced a relative deficit for abstract words, while lesioning the recurrent connections produced a relative deficit for concrete words. Standard neuropsychological reasoning would interpret this double dissociation as evidence of separate modules for abstract and concrete words, but there is no such modularization.
Similar problems arise in a developmental context. Different impairments imposed on the system at the beginning of training can differentially impact learning to perform correctly with items of different types. Again, standard neuropsychological reasoning would incorrectly imply that the underlying organization is modular.
To illustrate this point, we revisited Simulation One from TKS. We trained a single, three-layer, feed-forward network on both regular and exceptional past-tense forms using the same architecture and training patterns (kindly provided by Michael Thomas). Performance after training the intact network was 100% correct for items of both types, in line with the proposal (Rumelhart and McClelland, 1986) that a single integrated system might underlie the processing of both regular and exceptional forms1. We replicated their 'intact' condition, which produced 100% correct performance on both regulars and exceptions, and a fairly severe (80%) 'starting-point' lesion, which produced fairly good performance on regulars but poor performance on exceptions. We included another condition expected to differentially impair peformance on regulars (following Hoeffner and McClelland, 1993). This condition builds on the suggestion (Leonard et al, 1992; Leonard 1998) that the Englsih regular past tense inflection may be weakly represented in the speech signal, and therefore difficult to perceive for children with certain forms of language impairment (See also Tallal, 1995). Specifically, the non-syllabic forms of the English past tense (/t/ and /d/, as in 'liked' and 'loved') involve very slight additions that agree in voicing with the preceding sound, and can be very difficult to detect (Bird et al., in press), and the syllabic form (/^d/ as in 'hated') is unstressed. To simulate a deficit in perceiving these inflections, the units representing the past tense inflection were sometimes set to 0 in the target past-tense pattern that the network is given as a model for what it should learn. This is based on the idea that children learn from what they hear, and that the perceptual impairment makes the inflection sometimes imperceptible. This condition was otherwise identical to the intact condition. What we see in Table 1 is that in this new condition, there a disproportionate deficit in processing regular past tenses2.
Looking at Table 1, we clearly see a pattern of double dissociation. If we found two groups of children who exhibited the two patterns seen in the table and then employed standard neuropsychological reasoning, we would conclude that the normal brain contains separate systems for processing regular items and exceptions. This inference would be incorrect, however, since in this case we know that intact performance is generated by a single system that processes both regular and exceptional forms correctly. We simply have two different deficits that differentially impair learn to process the different types of items. A lesion that produces sparse connectivity reduces the ability of the network to become sensitive to particular combinations of input phonemes that must be considered simultaneously to correctly inflect an exception. An impairment that impacts perception reduces the network's exposure to the information that indicates the correct pronunciation of the regular past tense.
Table 1: Performance inflecting regular and exception words after training under two different forms of developmental abnormality.
| Deficit Type | Item Type | |
| Regular | Exception | |
| Intact Network | 100% | 100% |
| Sparse connectivity (80% of connections removed) |
75% | 22% |
| Perceptual Deficit (60% failure to perceive the regular inflection) |
38% | 97% |
In summary, Thomas & Karmiloff-Smith have sounded an important note of caution, indicating that standard neuropsychological reasoning cannot be applied uncritically to the interpretation of patterns of deficits seen in developmental disorders. We hope we have underscored their point by noting that this caution is important in interpreting adult as well as developmental cases. Our simulation suggests that developmental 'double dissociations' can be especially misleading, since developmental disorders can produce contrasting differential deficits in a single integrated mechanism, but the application of standard neuropsychological reasoning would interpret this pattern as evidence for a two-part system.
1. The simulation reported by TKS showed relatively poor generalization to novel items. There are several possible reasons for this: (1) the training corpus (which is based on one used previously by Plunkett and Marchman, 1993) employed a relatively large number of exceptions compared to regular forms; and (2) the patterns used for the variants of the regular past tense inflection do not reflect its phonological characteristics or its systematic relation to the phonological features of the stem. We would expect that a corpus that more realistically reflected the frequency structure of the language and the phonology of the regular inflection would produce a higher level of generalization.
2. We do not wish to defend the particulars of this simulation as an adequate model of the phonological impairment of any real children; specifically, we believe that in reality, such impairments affect perception of some aspects of exception items as well as regular items, so that the differential would not be as extreme.
Hoeffner, J., & McClelland, J. L. (1993). Can a perceptual processing deficit explain the impairment of inflectional morphology in developmental dysphasia? A computational investigation. In E. Clark (Ed.), Proceedings of the Twenty-fifth Annual Child Language Research Forum. Center for the Study of Language and Information, Stanford University, Palo Alto, CA. 38-49.
Joanisse, M.F. & Seidenberg, M. S. (1999). Impairments in verb morphology following brain injury: a connectionist model. Proceedings of the National Academy of Sciences (USA), 96, 7592-7597. Leonard, L. (1998). Children with specific language impairment. Cambridge, MA: MIT Press.
Leonard, L.B., McGregor, K.K., & Allen, G.D. (1992). Grammatical morphology and speech perception in children with specific language impairment. Journal of Speech and Hearing Research, 35, 1076-1085.
Pinker, S. (1991) Rules of language. Science, 253, 530-535.
Pinker & Ullman (2002). The past and Future of the Past Tense. Trends in Cognitive Sciences, 6, 456-463.
Plaut, D. C. (1995). Double dissociation without modularity: Evidence from connectionist neuropsychology. Journal of Clinical and Experimental Neuropsychology, 17, 291-321.
Plunkett, K. & Marchman, V. (1993). From rote learning to system building: Acquiring verb morphology in children and connectionist nets. Cognition 48, 21-69.
Rumelhart D. E., & McClelland J. L. (1986). On learning past tenses of English verbs. In McClelland, J.L. and Rumelhart, D. E. (Eds), Parallel Distributed Processing: Vol 2: Psychological and Biological Models. Cambridge, MA: MIT press.
Rumelhart, D. E., McClelland, J. L. & the PDP Research Group (1986). Parallel distributed processing: Explorations in the microstructure of cognition. Vols I and II. Cambridge, MA: MIT Press.
Shallice, T. (1988). From neuropsychology to mental structure. Cambridge, UK: Cambridge University Press.
Tallal, P. (1995). Development and disorders of speech and language: Implications for neural and behavioral plasticity. In B. Julesz and I. Kovacs (Eds.), Maturational windows and adult cortical plasticity. Reading, MA: Addison-Wesley, 239-246.
Thomas & Karmiloff-Smith (2002). Are developmental disorders like cases of adult brain damage? Implications from connectionist modelling. Behavioral and Brain Sciences.
Warrington, E. K. (1981). Concrete word dyslexia. British Journal of Psychology, 72, 175-196.
Warrington, E. K., & McCarthy, R. (1987). Categories of knowledge: Further fractionation and an attempted integration. Brain, 110, 1273-1296.