Hydrocephalus and Intelligence: The Hollow Men

Some claim the disease hydrocephalus reduces brain size by 95% but often with normal or even above-average intelligence, and thus brains aren’t really necessary. Neither is true.
statistics, psychology, IQ, biology
2015-07-282020-05-13 finished certainty: highly likely importance: 3


Hy­dro­cephalus is a dam­ag­ing brain dis­or­der where flu­ids com­press the brain, some­times dras­ti­cally de­creas­ing its vol­ume. While often ex­tremely harm­ful or life-threat­en­ing when un­treat­ed, some peo­ple with se­vere com­pres­sion nev­er­the­less are rel­a­tively nor­mal, and in one case (Lor­ber) they have been claimed to have IQs as high as 126 with a brain vol­ume 5% of nor­mal brains. A few of these case stud­ies have been used to ar­gue the ex­tra­or­di­nary claim that brain vol­ume has lit­tle or noth­ing to do with in­tel­li­gence; au­thors have ar­gued that hy­dro­cephalus sug­gests enor­mous un­tapped cog­ni­tive po­ten­tial which are tapped into rarely for re­pairs and can boost in­tel­li­gence on net, or that in­tel­li­gence/­con­scious­ness are non-ma­te­r­ial or tap­ping into ESP.

I point out why this claim is al­most cer­tainly un­true be­cause it pre­dicts count­less phe­nom­ena we never ob­serve, and in­ves­ti­gate the claimed ex­am­ples in more de­tail: the cases turn out to be sus­pi­ciously un­ver­i­fi­able (Lor­ber), likely fraud­u­lent (O­liveira), or ac­tu­ally low in­tel­li­gence (Feuil­let). It is un­clear if high­-func­tion­ing cases of hy­dro­cephalus even have less brain mass, as op­posed to lower proxy mea­sures like brain vol­ume.

I then sum­ma­rize an­thro­pol­o­gist John Hawk­s’s crit­i­cisms of the orig­i­nal hy­dro­cephalus au­thor: his brain imag­ing data could not have been as pre­cise as claimed, he stud­ied a se­lec­tive sam­ple, the story of the leg­endary IQ 126 hy­dro­cephalus pa­tient raises ques­tions as to how nor­mal or in­tel­li­gent he re­ally was, and hy­dro­cephalus in gen­eral ap­pears to be no more anom­alous or hard-to-ex­plain than many other kinds of brain in­juries, and in a com­par­ison, hemi­spherec­tomies, re­mov­ing or sev­er­ing a hemi­sphere, has pro­duced no anom­alous re­ports of above-av­er­age in­tel­li­gence (just deficit­s), though they ought to be just the same in terms of re­pairs or ESP.

That hy­dro­cephalus cases can reach roughly nor­mal lev­els of func­tion­ing, var­i­ous deficits aside, can be ex­plained by brain size be­ing one of sev­eral rel­e­vant vari­ables, brain plas­tic­ity en­abling cog­ni­tive flex­i­bil­ity & re­cov­ery from grad­u­al­ly-de­vel­op­ing con­di­tions, and over­pa­ra­me­ter­i­za­tion giv­ing ro­bust­ness to dam­age and poor en­vi­ron­ments, and learn­ing abil­i­ty. The field of deep learn­ing has ob­served sim­i­lar phe­nom­e­non in train­ing of ar­ti­fi­cial neural net­works. This is con­sis­tent with Lor­ber’s orig­i­nal con­tention that the brain was more ro­bust, and hy­dro­cephalus was more treat­able, than com­monly ac­cept­ed, but does not sup­port any of the more ex­otic in­ter­pre­ta­tions since put on his find­ings.

In short, there is lit­tle anom­alous to ex­plain, and stan­dard brain-cen­tric ac­counts ap­pear to ac­count for ex­ist­ing ver­i­fied ob­ser­va­tions with­out much prob­lem or re­sort to ex­tra­or­di­nary claims.

An ar­gu­ment re­cently sum­ma­rized by ex-bi­ol­o­gist SF au­thor ar­gues that in­tel­li­gence of hu­man brains may have lit­tle to do with brain size vari­ables such as neu­ron count, re­ly­ing on an ex­am­ple of , a med­ical con­di­tion:

For decades now, I have been haunted by the grainy, black­-and-white x-ray of a hu­man skull. It is alive but emp­ty, with a cav­ernous flu­id-filled space where the brain should be. A thin layer of brain tis­sue lines that cav­ity like an am­ni­otic sac. The im­age hails from a , the au­thor, re­ports that the pa­tient in ques­tion had “vir­tu­ally no brain”. But that’s not what scared me; hy­dro­cephalus is noth­ing new, and it takes more to creep out this ex-bi­ol­o­gist than a pic­ture of Ven­tri­cles Gone Wild.

The stuff of night­mares. (From )

What scared me was the fact that this vir­tu­ally brain-free pa­tient had an IQ of 126. He had a first-class hon­ors de­gree in math­e­mat­ics. He pre­sented nor­mally along all so­cial and cog­ni­tive ax­es. He did­n’t even re­al­ize there was any­thing wrong with him un­til he went to the doc­tor for some un­re­lated mal­a­dy, only to be re­ferred to a spe­cial­ist be­cause his head seemed a bit too large.

…The au­thors ad­vo­cate re­search into “Com­pu­ta­tional mod­els such as the smal­l­-world and scale-free net­work”—net­works whose nodes are clus­tered into high­ly-in­ter­con­nected “cliques”, while the cliques them­selves are more sparsely con­nected one to an­oth­er. De Oliveira et al sug­gest that they hold the se­cret to the re­silience of the hy­dro­cephalic brain. Such net­works re­sult in “higher dy­nam­i­cal com­plex­i­ty, lower wiring costs, and re­silience to tis­sue in­sults.” This also seems rem­i­nis­cent of those iso­lated hy­per­-effi­cient mod­ules of autis­tic sa­vants, which is un­likely to be a co­in­ci­dence: net­works from so­cial to ge­netic to neural have all been de­scribed as “smal­l­-world”…The point, though, is that un­der the right con­di­tions, brain dam­age may para­dox­i­cally re­sult in brain en­hance­ment. Smal­l­-world, scale-free net­work­ing—­fo­cused, in­ten­si­fied, over­clocked—might tur­bocharge a frag­ment of a brain into act­ing like the whole thing. Can you imag­ine what would hap­pen if we ap­plied that trick to a nor­mal brain?

Big if true. This is cer­tainly big if true: were 90%+ of the brain un­nec­es­sary, what could it do if the rest of it was used as effi­ciently as the small per­cent­age do­ing the ac­tual work? And this over­throws the en­tire un­der­stand­ing of what the brain is, how it func­tions, and how in­tel­li­gence works. Not to men­tion it has enor­mous im­pli­ca­tions for deep learn­ing: why do we throw these vast re­sources into train­ing ever-larger mod­els in or­der to in­crease their ca­pa­bil­i­ties, if all those neu­rons ac­tu­ally aren’t do­ing any­thing? We could boost AI ca­pa­bil­i­ties dra­mat­i­cally by dis­pens­ing with the re­dun­dan­cy.

Implausible Implications

These claims how­ever run into a lot of ob­jec­tions:

  1. All the ex­ist­ing ev­i­dence that Think­ing Is What Neu­rons Do: If neu­ron counts and cor­re­lates like brain vol­umes are so ir­rel­e­vant, why does point to their close cor­re­la­tion & causal­i­ty, from cross-species com­par­isons to hu­man ge­netic stud­ies (eg / & )? How do we ex­plain le­sion stud­ies or trau­matic brain in­juries or…

  2. Evo­lu­tion­ar­ily Ab­surd: if most of the brain can be dis­pensed with, then, as one of the most, if not the most, meta­bol­i­cal­ly-ex­pen­sive tis­sues in the body, why does evo­lu­tion not ex­tremely ag­gres­sively prune brains in hu­mans?

    And if it works in hu­mans, then why not every­where in the nat­ural world? (Where are the hy­dro­cephalus pri­mates, or birds, or any mam­mals? Where are the sib­ling species where one has 20x the neu­rons of the oth­er, be­cause the lat­ter has ‘flipped the switch’ and got­ten the hy­dro­cephalus ben­e­fit?) Differ­ences in to­tal vol­ume or brain den­sity are one thing, but a neu­ron is a neu­ron and will al­ways be ex­pen­sive to build, main­tain, and op­er­ate—if 5% works fine or bet­ter, why an ex­trav­a­gance of 20x?

  3. Fer­mi’s prob­lem: “Where Are The Aliens”? if the brain can work around such dam­age to the ex­tent of boost­ing in­tel­li­gence by 2SDs eas­ily while hav­ing only a frac­tion of the brain to work with, why do we not see this workaround trig­gered by any drug, surgery, or en­vi­ron­men­tal in­ter­ven­tion, or trig­gered spon­ta­neously at ran­dom, yield­ing po­ten­tially any­where up to 40S­Ds? We live in a big world with bil­lions of hu­mans, and any­thing that can hap­pen will hap­pen ()—ev­ery­thing from peo­ple to peo­ple . Hy­dro­cephalus is not that rare, and this mech­a­nism must not be rare ei­ther if such ex­treme cases can be found so ap­par­ently eas­i­ly. There should not be only a sin­gle for­get­table case.

    For per­spec­tive on what an ab­solute change of 20x neu­rons might look like, ; why do we not see any hu­mans where the gap be­tween their in­tel­li­gence and ours is many times larger than that be­tween us and chim­panzees?

  4. Where Are The Ge­niuses? Sim­i­lar­ly, sup­pos­ing that the re­pair mech­a­nism can never kick in for healthy hu­mans, re­sult­ing in su­per­in­tel­li­gences, why does the sup­posed ex­cess re­ported by Lor­ber not show up any­where? If hy­dro­cephalus can ac­tu­ally boost in­tel­li­gence on av­er­age, why do we not see gross over-rep­re­sen­ta­tions among high­-IQ co­horts or pro­fes­sions or sta­tus­es?

    It is a ba­sic point of or­der sta­tis­tics & tail effects that a small in­crease in a mean or vari­ance com­pared to a ref­er­ence group means that over­rep­re­sen­ta­tion will in­crease the more ex­treme the cas­es; even if we do not see su­per-hu­mans, we should still be struck by all the elite re­searchers who turn out to be hy­dro­cephalic (and the more elite they are, the worse their hy­dro­cephalus should be). But at least as far as Wikipedia is con­cerned, there are few hy­dro­cephal­ics of note for any rea­son.

Ex­tra­or­di­nary claims re­quire ex­tra­or­di­nary ev­i­dence. could se­ri­ously per­suade us that every­thing we know about the brain is wrong and that we need to re­boot neu­ro­science, as en­thus­es, to study how mem­ory and in­tel­li­gence are ac­tu­ally stored & con­ducted

in some ex­tremely min­ute, sub­atomic, form, as yet un­known to bio­chemists and phys­i­ol­o­gist­s…out­side the body—ex­tra­cor­po­re­al!…the brain [is] as a re­cep­tor/­trans­mit­ter of some form of elec­tro­mag­netic wave/­par­ti­cle… of course, when speak­ing of ex­tra­cor­po­real mem­ory we en­ter the do­main of “mind” or “spirit” with cor­re­spond­ing meta­phys­i­cal im­pli­ca­tions.

Problems With The Case Studies

Fake da­ta. For­tu­nately for Watts’s sleep, the case for hy­dro­cephalus is much worse than it looks. The brain scan he posts is not, in fact, of the IQ 126 case; Oliveira cap­tions it as im­ages from his lab of a nor­mal per­son­/nor­mal brain, a nor­mal per­son­/hy­dro­cephalus, and a hy­dro­cephalus pa­tient with “deep cog­ni­tive and mo­tor im­pair­ments”. Fur­ther, Oliveira et al lied about the ori­gin of the im­ages, which were copied from else­where, and the pa­per has been for­mally re­tract­ed.1

Sus­pi­ciously un­ver­i­fi­able anec­dotes. OK, but what about the Lewin ‘re­view’ (which is re­ally more of a jour­nal­is­tic ar­ti­cle) and the IQ 126 guy? Lewin pro­vides a pair of brain scans as well, but it’s im­por­tant to note that he never im­plies that that scan had any­thing to do with that pa­tient, and is retelling Lor­ber’s anec­dote at third hand. Lor­ber pro­vides no con­crete de­tails about him, in­clud­ing such ba­sics as what sort of IQ test and when it was done (mak­ing it po­ten­tially as mis­lead­ing as the peren­nial claim that Richard Feyn­man had a sim­i­lar IQ, based on an anec­dote in his bi­og­ra­phy about be­ing tested as a child). Sur­pris­ing­ly, as far as I can tell, Lor­ber and as­so­ciates (Lor­ber died 16 years lat­er, in 1996) have not pub­lished any­thing at all on their dataset in the 39 years since the Lewin press cov­er­age, and so there are no scans or data on this guy, or any of the oth­ers Lor­ber claimed to have above-nor­mal in­tel­li­gence.2

Mis­lead­ingly de­scribed anec­dotes. How many se­vere hy­dro­cephalus cases have even above-av­er­age in­tel­li­gence? De­spite Lor­ber’s claim to have found many cases of nor­mal and a highly above-av­er­age hy­dro­cephalus pa­tient eas­i­ly, sub­se­quent re­searchers ap­pear to have failed to do like­wise over the en­su­ing 4 decades. , cited by Watts, him­self only cites 3 in­stances: Lor­ber’s un­ver­i­fi­able anec­dotes via Lewin, the re­tracted & likely fraud­u­lent Oliveira et al, and a third case study of who re­ports their pa­tient hav­ing an IQ of 75.3 (I am puz­zled how Forsy­dke could pos­si­bly de­scribe Feuil­let’s IQ of 75, nearly re­tard­ed, as one of “two in­de­pen­dent con­fir­ma­tions” of Lor­ber, who was claim­ing a case with an IQ 3.4SDs higher than that; he says the same thing in Forsy­dke 2015 in crit­i­ciz­ing Hawk­s’s crit­i­cism, say­ing that Lor­ber had been vin­di­cated by Oliveira & Feuil­let. Forsy­dke oth­er­wise dodges all of Hawk­s’s points.)

In­creas­ingly less than meets the eye. So there ap­pear to be few ver­i­fi­able in­stances of se­vere hy­dro­cephal­ics with above-av­er­age, much less highly above-av­er­age, in­tel­li­gence. (I also have to won­der what more in­-depth test­ing of cas­es, go­ing be­yond sum­mary or ful­l-s­cale IQ tests to many other mea­sures of cog­ni­tive func­tion like com­plex re­ac­tion time, and tests over a lifes­pan, would show; one can only de­tect deficits in what & when one tests, after al­l.)

But let us take Lor­ber at face-value and ask, does a brain scan show­ing, say, 95% of vol­ume taken up by fluid im­ply that they have only 5% of the brain as a nor­mal per­son or only 5% of the neu­rons?

Hawks on Lorber

An­thro­pol­o­gist con­sid­ers what hy­dro­cephalus means and the Lor­ber anec­dote in 2007. He points out that:

  1. Mea­sure­ment Er­ror: 1970s CT scans were ex­tremely in­ac­cu­rate; it can­not de­tect ac­cu­rately at the 1mm as claimed, and can­not show that the cor­tex was uni­formly only 1mm, which is not that far from a nor­mal hu­man’s 2mm thick­ness. There may well be enough white mat­ter to­tal to still con­nect up the grey mat­ter. As Lowin noted in 1980, the in­ter­pre­ta­tion of brain scans is un­re­li­able and may not give a good es­ti­mate of what ex­actly is there and how much.

    Fur­ther, as few or no hy­dro­cephalus cases in the nor­mal range have been sub­ject to de­tailed scan­ning or post-mortem, it is un­clear if a given hy­dro­cephalus case in­volves any net loss of neu­rons. It is pos­si­ble that the high pres­sure merely in­creases neural den­si­ty, which ex­plains much of the cross-species differ­ences al­ready4 (One would not ex­pect this to be a good thing, but then, hy­dro­cephalus usu­ally is­n’t.)

  2. Ab­surd In­tel­li­gence Boosts: Lor­ber’s claimed sta­tis­tics im­ply in­creased in­tel­li­gence, not merely ro­bust­ness to dam­age:

    But the no­tion that “half” of the pa­tients where ven­tri­cle ex­pan­sion is greater than 95% of the cra­nium have IQ’s greater than 100 is math­e­mat­i­cally im­plau­si­ble. The de­fi­n­i­tion of IQ is that the mean is 100. This means that only half of peo­ple with­out ven­tri­cle ex­pan­sion have IQ over 100. Lor­ber seems to have claimed that the most se­vere cases of hy­dro­cephalus ac­tu­ally see an in­crease in the pro­por­tion of high­-IQ in­di­vid­u­als, de­spite “many” be­ing se­verely dis­abled. I’m not say­ing it’s im­pos­si­ble, but like “a mil­lime­ter or so,” this is the kind of sta­tis­tic that de­serves skep­ti­cism.

  3. Se­lec­tion Bias: Lor­ber was highly se­lec­tive in the cases he stud­ied, pre­cisely be­cause he was look­ing for anom­alies; such cases in­trin­si­cally will have un­known com­pen­sat­ing fac­tors (such as high ge­netic po­ten­tial for in­tel­li­gence, which can buffer the dam­age), and can yield highly mis­lead­ing re­sults through things like

  4. Miss­ing Caveats: the story of the IQ 126 stu­dent is bizarre, un­less Lor­ber is leav­ing some­thing out­—who is re­ferred to a hy­dro­cephalus spe­cial­ist for ex­pen­sive & un­usual brain imag­ing scans (e­spe­cially back then) just be­cause “his head seemed a bit too large”?

    Nor is it ob­vi­ous from the re­ports that the con­di­tion had “no” cog­ni­tive man­i­fes­ta­tions. Much seems to de­pend on the sin­gle case de­scribed above, with an ap­par­ently nor­mal col­lege stu­dent walk­ing in off the street to dis­cover he had min­i­mal brain mass. But this story is quite ob­vi­ously in­cred­i­ble as pre­sent­ed: most neu­rol­o­gists don’t per­form brain scans just be­cause a col­lege stu­dent wears a large hat. It seems rea­son­able to in­fer that the stu­dent was re­ferred by his doc­tor to Lor­ber, a hy­dro­cephalus spe­cial­ist, for some rea­son. We can only guess what the rea­son might be, but it hardly gives con­fi­dence in the anec­dote!

  5. Ques­tion­able Dam­age: Hawks con­cludes that anom­alies may not be too use­ful to study here:

    With­out ques­tion, there are many pa­tients who have this out­come—no sig­nifi­cant cog­ni­tive deficit com­pared to non­pa­tients, de­spite pro­found pathol­o­gy. This is true of al­most any pathol­ogy affect­ing the brain, in­clud­ing tu­mors, strokes, and de­vel­op­men­tal ab­nor­mal­i­ties. The ques­tion is whether this pro­vides a valid model for un­der­stand­ing the adap­tive im­por­tance of brain vol­ume. It seems that later on­set hy­dro­cephalus, where a nor­mal brain is com­pressed within a rel­a­tively nor­mal-sized skull by cere­brospinal fluid pres­sure, does not re­ally ap­ply to the evo­lu­tion­ary ques­tion. The re­ported cases do not ap­par­ently in­volve sig­nifi­cant gray mat­ter tis­sue loss. A “thin” cor­tex does not nec­es­sar­ily im­ply func­tion­ally small cor­ti­cal vol­ume, even with sub­stan­tial white tis­sue loss.

Coun­ter-ex­am­ple: no high­-IQ hemi­spherec­tomies. Hawks then goes on to ex­am­ine , an op­er­a­tion to treat the most se­vere epilep­tic cases by sev­er­ing one brain hemi­sphere from an­other or re­mov­ing it en­tire­ly. Un­sur­pris­ing­ly, hemi­spherec­tomy pa­tients have many se­vere cog­ni­tive prob­lems: note, among their many lan­guage/read­ing/be­hav­ioral/school prob­lems re­quir­ing in­ten­sive in­ter­ven­tions such as spe­cial school­ing (as well as evo­lu­tion­ar­i­ly-rel­e­vant side-effects like par­tial paral­y­sis), that only “5 (21%) of the 24 pa­tients older than 18 years of age were gain­fully em­ployed.” This is still a suc­cess for hemi­spherec­to­my, since se­vere epilepsy badly dam­ages nor­mal de­vel­op­ment and can be fa­tal (as Hawks points out, since the brain tis­sue be­ing re­moved or sev­ered is so badly dam­aged al­ready, it’s not nec­es­sar­ily much of a loss), and in­ci­den­tally demon­strates how / al­lows com­pen­sa­tion for los­ing ac­cess to a whole hemi­sphere, at least when done young enough. An­other ex­treme case is : the true loss of brain, un­like hy­dro­cephalus, but there are no case re­ports of nor­mal or above-av­er­age cog­ni­tive func­tion­ing; hy­dra­nen­cephaly cas­es, while nor­mal-seem­ing for months after birth, are in­stead pro­foundly dis­abled, barely re­spond to stim­uli, and are lucky to live for more than a few years (), with 19 years be­ing the record (the case).

What Does Hydrocephalus Mean?

Hawks con­cludes:

So the case of hemi­spherec­tomy does not test the propo­si­tion that nor­mal cog­ni­tive per­for­mance is pos­si­ble after a great re­duc­tion in brain size. In­stead it pos­si­bly tests the propo­si­tion that a re­duc­tion in brain size may be con­sis­tent with nor­mal cog­ni­tive per­for­mance un­der a spe­cial­ized cul­tural and en­vi­ron­men­tal regime. That hy­poth­e­sis is re­futed by the ma­jor­ity of cases in the clin­i­cal record, for whom the spe­cial­ized learn­ing en­vi­ron­ment has not man­aged to elim­i­nate de­vel­op­men­tal deficits. Still, for many pa­tients some com­bi­na­tion of surgery, ther­apy and learn­ing as­sis­tance do make a de­ci­sive differ­ence, and they at­tain nor­mal cog­ni­tive per­for­mance—even nor­mal for de­vel­op­men­tal age.

How do these cases ap­ply?

There is no sin­gle con­clu­sion that we can draw from these ex­am­ples of ex­treme patho­log­i­cal re­duc­tion in brain size in hu­mans. Clear­ly, the brain is ca­pa­ble of re­mark­able plas­tic­ity in de­vel­op­ment, in­clud­ing al­ter­nate lo­cal­iza­tions of some func­tions that are highly lo­cal­ized in most adults. But can we ap­ply this plas­tic­ity more gen­er­al­ly, to sug­gest that al­most any brain struc­ture might have evolved in an­cient hu­man pop­u­la­tions? Even those that in­volve im­mense re­duc­tions in over­all brain size?

Ro­bust­ness to dam­age & en­vi­ron­men­tal sub­op­ti­mal­i­ty. It should be men­tioned that these as­sess­ments build on a rather nar­row view of “cog­ni­tion.” For in­stance, all hemi­spherec­tomy pa­tients have some paral­y­sis on the op­po­site side from the re­moved hemi­sphere. The func­tions of the mo­tor and sen­sory cor­tices of the ab­sent side do not ap­pear to have the de­vel­op­men­tal plas­tic­ity ex­hib­ited by lan­guage. From the per­spec­tive of fit­ness in pre­his­toric hu­man pop­u­la­tions, the ad­e­quate con­trol of move­ment and per­cep­tion of sen­sory in­for­ma­tion would have a sub­stan­tially greater im­por­tance than in to­day’s cul­tural mi­lieu. So a re­duc­tion in brain size that im­pacts mo­tor and sen­sory func­tion but leaves other as­pects of cog­ni­tion in­tact cer­tainly can­not be said to have no im­pact. Just be­cause a re­duc­tion in per­for­mance can be man­aged within our pop­u­la­tion does not mean that it could have evolved in some past pop­u­la­tion. Al­so, the at­tain­ment of “nor­mal” cog­ni­tion, how­ever de­fined, re­quires sub­stan­tial in­vest­ment and teach­ing for the av­er­age hu­man. Hu­mans with de­vel­op­men­tal chal­lenges often can at­tain nor­mal cog­ni­tive per­for­mance for their age, par­tic­u­larly when sup­ple­men­tary teach­ing and ther­apy is avail­able. All this is to say that hu­man brains are coad­apted with be­hav­ioral pat­terns that chan­nel de­vel­op­men­t…Although it may be con­ceiv­able—even if it is far from demon­strat­ed—that a rad­i­cally smaller brain cou­pled with a spe­cial­ized cul­ture might have in­creased fit­ness, ap­par­ently there was no avail­able evo­lu­tion­ary path­way to that adap­ta­tion. I would guess that it is sim­ply more diffi­cult to main­tain the nec­es­sary cul­tural spe­cial­iza­tions for such an adap­ta­tion within the con­text of an­cient hu­man pop­u­la­tion struc­ture. It is eas­ier to ac­com­plish de­vel­op­ment with a large brain that can em­ploy many bot­tom-up strate­gies to build its cog­ni­tive abil­i­ties.

Thus the lost (or just com­pressed) white mat­ter may be no more ‘un­nec­es­sary’ than, say, one’s spare lung or kid­ney, most of one’s liv­er, or any of one’s limbs… , de­scrib­ing a mouse hy­dro­cephalus case, ask:

Sur­vival re­quires sens­ing the en­vi­ron­ment, pro­cess­ing the in­for­ma­tion, and re­spond­ing ap­pro­pri­ate­ly. There is no mor­tal threat or com­pe­ti­tion to sur­vive for a rat main­tained in a cage for two years un­der con­trolled en­vi­ron­men­tal con­di­tions with food and wa­ter ad li­bi­tum. R222 is a study of the fun­da­men­tal neu­ro­bi­o­log­i­cal and be­hav­ioral processes that sus­tain a re­source ad­e­quate, safe, am­bu­la­tory life. The sever­ity of the hy­dro­cephalus in R222 in the face of nor­mal body weight and growth, nor­mal mo­tor be­hav­ior and spa­tial mem­o­ry, and evoked ac­tiv­ity to smells, tac­tile stim­u­la­tion and vi­sion would sug­gest neu­road­ap­ta­tion to a life-long ab­nor­mal­i­ty. This rare case can be viewed as one of na­ture’s mir­a­cles pro­vid­ing the unique op­por­tu­nity to ex­am­ine the brain’s ca­pac­ity for neu­ro­plas­tic­ity and re­or­ga­ni­za­tion nec­es­sary for sur­vival…Is the cor­tex nec­es­sary? There have been nu­mer­ous stud­ies across a va­ri­ety of mam­mals look­ing at the de­vel­op­men­tal con­se­quences of rad­i­cal decor­ti­ca­tion in neonates24,25,26,27,28,29,30. While there are mi­nor deficits par­tic­u­larly in some mo­tor pat­terns and mo­tor co­or­di­na­tion, the decor­ti­cate an­i­mal can eat, drink, sleep and grow to nor­mal size. They re­spond to vi­sual and au­di­tory stim­uli. They dis­play nor­mal species-spe­cific so­cial, ma­ter­nal, ag­gres­sive and sex­ual be­hav­iors. They mate and re­pro­duce. Na­ture again has pro­vided sci­ence with an ex­treme form of decor­ti­ca­tion in hu­man­s–hy­dra­nen­cepha­ly, a rare, in­her­ited dis­or­der where by ba­bies are born with­out cere­bral hemi­spheres. There is no treat­ment, yet in­cred­i­bly, with the proper care and sta­bi­liza­tion these in­di­vid­u­als can live for years31, not in a veg­e­ta­tive state and are re­spon­sive to their sur­round­ings.

This might sound rea­son­able, but how does Fer­ris de­fine ‘nec­es­sary’? Hu­man hy­dra­nen­cephaly cas­es, as dis­cussed, are lit­tle bet­ter off than coma vic­tims: McAbee de­scribe their first case’s “re­spon­sive to their sur­round­ings” as con­sist­ing of “turn[ing] his head ip­si­lat­er­ally to sounds, mu­sic, and venipunc­ture.” & re­quir­ing a feed­ing tube, and the sec­ond case “He had no defi­nite aware­ness of his en­vi­ron­ment. The pupil­lary re­sponse to light was min­i­mal; he did not fix or fol­low vi­su­ally and had no vi­sual re­sponse to threat.” How about the mat­ing and re­pro­duc­ing—­surely com­plex be­hav­iors, and if decor­ti­ca­tion al­lows even that, what do we need a cor­tex for? Look­ing at Fer­ris et al 2019’s ci­ta­tions, the only two ref­er­ences for that are less than im­pres­sive: for fe­males (), “mate and re­pro­duce” ap­par­ently means only pas­sively lay­ing in place with one’s rump el­e­vated in , and for the males (), when paired with re­cep­tive (nor­mal) fe­males, they were more in­fer­tile than nor­mal rats and had to be man­u­ally “stim­u­lated” by the re­searchers! These cases show that the cor­tex is un­nec­es­sary only if one has a low stan­dard for what is nec­es­sary, which is de­tached from any re­al-world con­sid­er­a­tions.

I would pre­dict that deeper study of such hy­dro­cephalus cases would re­veal the fol­low­ing:

  1. far less neural loss than ex­pected based on vol­ume changes, con­sis­tent with loss of a large per­cent­age be­ing crip­pling or fa­tal

  2. a cor­re­la­tion be­tween neu­ron loss & per­for­mance deficits, where high­er-per­form­ing hy­dro­cephalus is due to greater den­sity for a given vol­ume re­duc­tion or the same mass be­ing dis­trib­uted in a weirder and hard­er-to-mea­sure way

  3. sys­temic per­for­mance deficits, where small deficits in each task add up to large to­tal deficits on com­plex tasks—in­creas­ing in harder more-nat­u­ral­is­tic tasks re­quir­ing more mem­o­ry, learn­ing, and cog­ni­tive flex­i­bil­ity

  4. re­pro­duc­tive fit­ness penal­ties, with greater fit­ness penal­ties in en­vi­ron­ments which are more com­pet­i­tive & re­sem­ble the an­ces­tral en­vi­ron­ment (eg the ), as op­posed to mod­ern en­vi­ron­ments which may be un­de­mand­ing.

    Con­sid­er­ing cases like Feuil­let’s case, who worked as a French “civil ser­vant”, one might won­der if this in­di­cates in­tel­li­gence is not all that im­por­tant: how could some­one with an IQ of 74 so much as feed and dress them­selves, one might say, much less hold a “white-col­lar job” or be mar­ried with chil­dren?

    But one should re­mem­ber that cases of men­tal re­tar­da­tion may be caused by dis­or­ders that have many pro­foundly dis­abling side-effects, ex­ag­ger­at­ing the dys­func­tion­al­i­ty; and that oth­er­wise-healthy nor­mal in­di­vid­u­als can have many cop­ing strate­gies & sup­port­ive en­vi­ron­ments even while still hav­ing as­ton­ish­ing deficits. This is com­pa­ra­ble to cog­ni­tive de­cline & se­nil­ity in the el­der­ly, or to adult il­lit­er­a­cy: il­lit­er­ate adults (often due to dyslex­ia) can be sur­pris­ingly func­tional & fool every­one who knows them by ex­ploit­ing a range of tricks, rang­ing from heavy re­liance on oral com­mu­ni­ca­tion like tele­phones and get­ting other peo­ple to read for them to avoid­ing writ­ing-heavy en­vi­ron­ments in fa­vor of con­crete ones to skilled guess­ing to mem­o­riz­ing spe­cific pieces of writ­ing & pre­tend­ing to read it to sim­ply ly­ing. Hu­man Rights Watch 2001 de­scribes how re­tarded crim­i­nal de­fen­dants, even ones be­ing pros­e­cuted , would deny they are re­tarded or fool their lawyers into sign­ing them up for col­lege cal­cu­lus or lied about fin­ish­ing high school.

    Nev­er­the­less, de­spite the il­lu­sion of com­pe­ten­cy, the re­tarded may still have pro­found deficits, as Hu­man Rights Watch 2001 then goes on to give ex­am­ples: de­fen­dants who don’t know how to use stamps or what their stom­ach does, who count on their hands, who will agree to any­thing an au­thor­ity fig­ure says, who don’t know why rape is wrong (sug­gest­ing be­cause they lacked “per­mis­sion” or it was “against her re­li­gion”), who un­der­stand stab­bing but not shoot­ing kills some­one, who smile or sleep through their tri­als, who ask what to wear to their fu­neral after be­ing ex­e­cut­ed… In or­di­nary cir­cum­stances, which per­mit ex­treme lev­els of rep­e­ti­tion & re­liance on oth­ers to sup­port a safe en­vi­ron­ment, they may ap­pear as func­tional as any­one else, yet they will make rare (but ex­tremely im­por­tant) er­rors and can­not han­dle novel prob­lems.5 (To ruin one’s life, one only has to screw up once.) Gre­go­ry’s Mc­Na­ma­ra’s Fol­ly: The Use of Low-IQ Troops in the Viet­nam War de­scribes many ex­am­ples from of re­tarded or low-in­tel­li­gence men who could func­tion in a safe civil­ian en­vi­ron­ment do­ing sim­pli­fied me­nial labor, but could not be trained to do any­thing safely in the mil­i­tary or cope with novel en­vi­ron­ments, suffer­ing ex­treme ca­su­alties, often self­-in­flict­ed. While any in­di­vid­ual in­stant may seem unim­por­tan­t—there are many ways to avoid need­ing to un­der­stand how to use stamps one­self, and most peo­ple would never rape a woman even if they don’t un­der­stand why rape is bad—they add up over a life­time and push the en­tire pop­u­la­tion to­wards bad out­comes. One might note that Feuil­let’s case had a be­low-av­er­age num­ber of chil­dren and his “white-col­lar” job may mean noth­ing more than stamp­ing the same form for 40 years in the DMV as a kind of covert wel­fare; or con­sider the pro­found con­se­quences for in­di­vid­u­als with re­duced/no pain sen­si­tiv­ity.

Analogies from Deep Learning

Ar­ti­fi­cial par­al­lels in Deep Learn­ing. Hawk­s’s dis­cus­sion of chan­nel­ing leads to an in­ter­est­ing par­al­lel with deep learn­ing (DL): it has been re­peat­edly es­tab­lished that all neural net­works are ‘over­pa­ra­me­ter­ized’ in the sense that what­ever per­for­mance a NN reaches at the end of train­ing, one can eas­ily ‘dis­till’ or ‘com­press’ it into a smaller (or faster) NN with slightly worse per­for­mance, by at least a frac­tion and some­times by or­ders of mag­ni­tudes.

De­spite neu­ron/­pa­ra­me­ter count clearly caus­ing NN ‘in­tel­li­gence’. Pa­ra­me­ter & neu­ron counts in NNs clearly do cause big­ger NNs are more ex­pres­sive than smaller ones, be­cause when we ex­per­i­men­tally vary neural net size, we see log­a­rith­mic/pow­er-law in­creases in per­for­mance, even for the largest train­able NNs; but at the same time, we still see that of course size does not ex­plain 100% of vari­ance in per­for­mance—the same ar­chi­tec­ture trained in the same way on the same data will yield NNs with no­tice­ably differ­ent fi­nal per­for­mance, sim­ply due to the sto­chasc­ity of the ini­tial­iza­tion, reg­u­lar­iza­tion mech­a­nisms like dropout, and data se­lec­tion6, and it would be pos­si­ble to find runs where a smaller NN out­per­forms a larger NN which is oth­er­wise ex­actly the same, yet, nev­er­the­less, larger is more pow­er­ful. Sim­i­lar­ly, dis­til­la­tion/­com­pres­sion work quite well, and yet can­not be tap­ping into any sort of du­al­ism or psi.

DL demon­strates func­tional ben­e­fits of over­pa­ra­me­ter­i­za­tion. Why? The re­cent work on the strongly in­di­cates that most of the work is be­ing done by a rel­a­tively small part of the net­work, and the ex­pres­sive power of the full mod­el, which is enor­mously rich, is largely un­nec­es­sary for the task we ap­ply it to. This raises the puz­zling ques­tion of how we can get away with train­ing such large mod­els with­out them over­fit­ting or learn­ing use­less things, when stan­dard sta­tis­ti­cal & ma­chine learn­ing the­ory teaches that this ought to be a car­di­nal sin and guar­an­teed fu­tile. And pe­cu­liar­ly, know­ing that our NN is far too big & we only need a small NN does not let one train the fi­nal smaller NN di­rect­ly: if one takes a smaller NN and trains a new from-scratch NN with the same size (which one knows is big enough by con­struc­tion), it sim­ply does­n’t work.

Over­pa­ra­me­ter­i­za­tion means al­ways pos­si­ble im­prove­ments. Why not? The an­swer seems to be that over­pa­ra­me­ter­i­za­tion is nec­es­sary to en­able effi­cient learn­ing, by mak­ing the ‘fit­ness land­scape’ smooth and en­sur­ing there is al­ways a pa­ra­me­ter which can be tweaked to make some learn­ing pro­gress, while smaller NNs are more likely to get ‘stuck’ in a bad lo­cal op­ti­mum, and have no way out. (An ex­am­ple of a ‘rough’ fit­ness land­scape would be al­most all com­puter pro­gram­ming lan­guages: delet­ing a sin­gle char­ac­ter like a semi­colon may make a pro­gram go from 100% per­for­mance on a task to 0% and per­haps may not even com­pile.) This is par­tially due to hav­ing so many ran­dom sub­-net­works that at least one of them ‘wins the lot­tery’ and is a good fit for the cur­rent task. Thus, one can start off with a big model and then learn effec­tive­ly, and spe­cial­ize down into a much smaller but in­flex­i­ble mod­el, but one can­not learn the smaller model to be­gin with from the raw data (un­less one is ex­ceed­ingly lucky and the small model starts off with a use­ful ini­tial­iza­tion and hap­pens to not run into any trap­s); the small model can only be cre­ated based on heavy guid­ance from the big mod­el.

For a hu­man brain, which must do life­long learn­ing on a con­stantly chang­ing mix of tasks, in a harsh evo­lu­tion­ary con­text filled with com­pet­i­tive ze­ro-sum dy­nam­ics where small differ­ences can mat­ter & there is no ‘big model’ to copy or learn from, this sug­gests that re­ly­ing on a small brain would be a ter­ri­ble idea. Since—un­like soft­ware mod­el­s—here is no way to train a hu­man brain to adult­hood and then scoop it out and shrink it to a much more effi­cient ver­sion and copy it every­where else, re­gard­less of how good it has be­come, every hu­man brain must start effec­tively from scratch, and can­not ben­e­fit from such com­pres­sion, and hu­man brains must be over­pa­ra­me­ter­ized to en­sure it can meet what­ever chal­lenges the world throws at it. (Soft­ware, on the other hand, can be copied, which has in­ter­est­ing im­pli­ca­tions for the ‘hard­ware over­hang’ ar­gu­men­t.)

Conclusion: No Evidence

Un­like­ly, poor ev­i­dence, and eas­ily ex­plained. So to sum up: peo­ple have claimed that hy­dro­cephalus de­stroys most of the brain yet hav­ing only a frac­tion of the brain is con­sis­tent with nor­mal or above-av­er­age in­tel­li­gence, and may even in­crease it. This is ex­tremely im­plau­si­ble based on every­thing we know about in­tel­li­gence and evo­lu­tion and pop­u­la­tion dis­tri­b­u­tions and is a bad de­scrip­tion of what hy­dro­cephalus does (con­flat­ing dis­tri­b­u­tion & vol­ume with brain mat­ter). On av­er­age, hy­dro­cephalus and sim­i­lar things like hemi­spherec­tomy do (as ex­pect­ed) in­duce many deficits rang­ing from mild to sev­ere, which can be ac­com­mo­dated to some de­gree by neural plas­tic­i­ty, in­di­vid­ual differ­ences, ex­ten­sive en­vi­ron­men­tal in­ter­ven­tions, and free­dom from nat­ural se­lec­tion. Claims to the con­trary, aside from fail­ing to deal with the many ob­jec­tions that this is im­plau­si­ble, turn out to be based on un­doc­u­ment­ed, fraud­u­lent, or mis­lead­ingly de­scribed cas­es, and pri­mar­ily pushed by cranks. Ul­ti­mate­ly, hy­dro­cephalus does not ap­pear to present any par­tic­u­lar chal­lenge to the stan­dard un­der­stand­ing of in­tel­li­gence as be­ing caused by a ma­te­r­ial brain whose effi­ciency at cog­ni­tive tasks is dri­ven by neu­ron count, wiring pat­terns, neural in­tegri­ty, and gen­eral health: not only is the ev­i­dence ex­tra­or­di­nar­ily in­ad­e­quate to jus­tify the ex­tra­or­di­nary claims made by some au­thors, it is un­clear how much, if any, ev­i­dence there at all.


  1. My notes say the pla­gia­rism was first spot­ted by Ondřej Havlíček on the Neu­roskep­tic blog. Why would any re­searcher lie about im­ages like that? Who knows. But it demon­strates why one can’t put too much trust in any sin­gle dat­a­point. I have raised my con­cerns about the im­plau­si­bil­ity of these hy­dro­cephalus claims & the re­trac­tion with Watts on his blog and 3 emails so far.↩︎

  2. I find such la­cu­nae to be sus­pi­cious. When you look at the most ques­tion­able find­ings and ex­per­i­ments in Re­pro­ducibil­ity Cri­sis & psy­chol­o­gy-re­lated mat­ters, like the , , or , one of the more com­mon hall­marks is a sin­gle ex­tra­or­di­nary find­ing which makes a splash with press at­ten­tion, and then a strange discli­na­tion by the orig­i­nal au­thor to repli­cate their own find­ing or pub­lish any­thing fur­ther on it which is not a sec­ondary-source-style re­hash.↩︎

  3. Feuil­let et al 2007 is some­times mis­quoted as mea­sur­ing an IQ of 84; that refers only to the high­est sub­test score, not the to­tal test score.↩︎

  4. See Her­cu­lano-Houzel’s re­search, like , Her­cu­lano-Houzel 2012.↩︎

  5. One might say that this is in­her­ently true of any kind of ab­strac­tion or con­cept or model or long-term mem­o­ry: if one needed them fre­quently, they could be han­dled di­rectly by tri­al-and-er­ror or short­-term mem­o­ry. If eval­u­ated in or­di­nary com­mon cir­cum­stances, their value will be hid­den. They are valu­able be­cause they can ex­trap­o­late to novel cir­cum­stances, or while used rarely and most uses of lit­tle val­ue, a few uses are ex­tremely im­por­tant. A fact may be use­ful just once in a life­time, but that use may save one’s life; and, in all the years up to the dan­ger, such a per­son would look iden­ti­cal to some­one who did­n’t or could­n’t learn that fact. One is re­minded of how neural net-gen­er­ated text like can look hu­man-writ­ten save for a sin­gle stray char­ac­ter—a “her” in­stead of “he”—which shat­ters the il­lu­sion.↩︎

  6. Or hard­ware non­de­ter­min­ism— and not guar­an­teed to re­turn the same re­sult for a given op­er­a­tion! This is just one of the many is­sues with ex­act re­pro­ducibil­ity in deep learn­ing. Deep re­in­force­ment learn­ing par­tic­u­larly strug­gles with this lack of re­pro­ducibil­ity and still higher vari­ance than reg­u­lar DL.↩︎