Simulating Brain DamageAdults with brain damage make some bizarre errors when readingwords. If a network of simulated neurons is trained to readand then is damaged, it produces strikingly similar behaviorby GeoÝrey E. Hinton, David C. Plaut and Tim ShalliceIn 1944 a young soldier suÝered abullet wound to the head. He sur-vived the war with a strange disabil-ity: although he could read and compre-hend some words with ease, many oth-ers gave him trouble. He read the wordantique as ÒvaseÓ and uncle as Ònephew.ÓThe injury was devastating to the pa-tient, G.R., but it provided invaluableinformation to researchers investigat-ing the mechanisms by which the braincomprehends written language. A prop-erly functioning system for convert-ing letters on a page to spoken sounds reveals little of its inner structure, butwhen that system is disrupted, the pecu-liar pattern of the resulting dysfunctionmay oÝer essential clues to the original,undamaged architecture.During the past few years, computersimulations of brain function have ad-vanced to the point where they can beused to model information-processingpathways. We have found that deliberatedamage to artiÞcial systems can mimicthe symptoms displayed by people whohave sustained brain injury. Indeed,building a model that makes the sameerrors as brain-injured people do givesus conÞdence that we are on the righttrack in trying to understand how thebrain works.We have yet to make computer mod-els that exhibit even a tiny fraction ofthe capabilities of the human brain. Nev-ertheless, our results so far have pro-duced unexpected insights into the waythe brain transforms a string of let-ter shapes into the meaning of a word.When John C. Marshall and FredaNewcombe of the University of Oxfordanalyzed G.R.Õs residual problems in1966, they found a highly idiosyncraticpattern of reading deÞcits. In additionto his many semantic errors, G.R. madevisual ones, reading stock as ÒshockÓand crowd as Òcrown.Ó Many of his mis-readings resembled the correct word inboth form and meaning; for example,he saw wise and said Òwisdom.ÓDetailed testing showed that G.R.could read concrete words, such as ta-ble, much more easily than abstractwords, such as truth. He was fair at read-ing nouns (46 percent correct), worseat adjectives (16 percent), still worse at verbs (6 percent) and worst of all at76 SCIENTIFIC AMERICAN October 1993BRAIN IMAGES show damage to the language-processing areasof patients with acquired dyslexia, which can now be mod-eled by artiÞcial neural networks. (These positron-emissiontomography scans, made by Cathy J. Price and her colleaguesat the MRC Cyclotron Unit in London, measure activity of thebrain in successive horizontal slices, starting at the top. LowCopyright 1993 Scientific American, Inc.function words, such as of (2 percent).Finally, he found it impossible to readwordlike nonsense letter strings, suchas mave or nust.Since then, clinicians have studiedmore than 50 other patients who makesemantic errors in reading aloud, andvirtually all of them show the samestrange combination of symptoms. In1973 Marshall and Newcombe describedtwo contrasting types of acquired dys-lexia. So-called surface dyslexics mis-read words that are pronounced in anunusual way, often giving the more reg-ular pronunciation; a surface dyslexicmight read yacht as Òyatched.Ó In con-trast, a ÒdeepÓ dyslexic patient like G.R.might read yacht as Òboat.ÓTo explain the existence of these two types of dyslexia, Marshall andNewcombe proposed that the informa-tion processed in normal reading trav-els along two distinct, complementaryroutes. Surface dyslexics retain the pho-nological route, which relies on com-mon spelling-to-sound correspondenc-es. Deep dyslexics, meanwhile, retainthe semantic route, which allows themeaning of a word to be derived direct-ly from its visual form (when it can bederived at all). A person reading wordsaloud via the semantic route derivespronunciation entirely from meaning.According to Marshall and Newcombe,the errors produced by deep dyslexicsreßect how the semantic route oper-ates in isolation. Later empirical Þnd-ings suggest that this account is over-simpliÞed, but the notion of a semanticroute is still generally accepted. It nowseems likely that deep dyslexics not onlylose their phonological route but havedamage somewhere along the semanticone as well.The hypothesis that reading de-pends on multiple routes that canbe separately damaged has provedfruitful in classifying patients but lessuseful in understanding the precise na-ture of their injuries. Max Coltheart ofMacquarie University in Australia andEleanor M. SaÝran of Temple Universityhave both proposed, for example, thatthe reading of deep dyslexics may beara strong resemblance to that of patientswho have only the right hemisphere oftheir brain functioning.This explanation, however, provideslittle insight into the highly characteris-tic pattern of errors that typically oc-curs in acquired dyslexia. Any detailedexplanation of how errors arise andwhy they form consistent patterns re-quires a model of how that informationis processed in each routeÑand of howthis processing goes wrong when theneural circuitry is damaged. Psycholo-gists often use abstract, algorithmic de-scriptions of the way that the brain han-dles information. These descriptions ob-viously cannot be subjected to the kindsof injuries that brain cells may incur.As a result, we have turned to neuralnetworksÑidealized computer simula-tions of ensembles of neurons. We havedeveloped networks that perform therole of the semantic route, and then wehave selectively removed connectionsbetween neurons to see how their be-GEOFFREY HINTON, DAVID PLAUT andTIM SHALLICE use artiÞcial neural net-works to investigate the behavior of thebrain. Hinton is the Noranda Fellow ofthe Canadian Institute for Advanced Re-search and professor of computer sci-ence and psychology at the University ofToronto. He has been studying represen-tation and learning in neural networksfor more than 20 years. Plaut is a post-doctoral research associate in the psy-chology department at Carnegie MellonUniversity, where he earned his doctor-ate in computer science in 1991. Shal-lice is professor of psychology at Univer-sity College, London, where he receivedhis doctorate in 1965. His research hasmainly focused on what can be under-stood about the normal cognitive sys-tem by studying impairments result-ing from neurological disease, with