F o u r n i s s e u r s & P a r t e n a i r e s R e l a t ...
R E F E R E N C E F R A M E S P IN G L A S S V II: S P IN ...
Transcript of R E F E R E N C E F R A M E S P IN G L A S S V II: S P IN ...
REFERENCE FRAME
SPIN GLASS VII: SPIN GLASSAS PARADIGMPhilip W. Anderson
In my last column on spin glasses(September 1989, page 9) I tried toshow you that the exact solutions of aparticular spin glass problem, byGiorgio Parisi and Gerard Toulouse,gave us great insight into the theoryof complex optimization problems, aswell as an algorithm for solving someof them. One such problem, whichhas been exhaustively studied bymethods of spin glass theory, is thegraph partition problem. This is thequestion of how to divide an arbitrarygraph into two pieces, cutting thefewest possible bonds. My studentYao-Tian Fu (now at WashingtonUniversity, St. Louis) initiated thestudy of the graph partition problemby replica theory. This classic prob-lem of complexity theory was difficultto solve for sparse graphs by thosemethods, but another of my students,Wuwell Liao, seems to have done it.
Even more interesting than theseapplications to complexity theory isthe way apparently unrelated areasof science have been stimulated intoparallel growth by the spin glasswork. John Hopfield (Caltech), whowas instrumental in bringing me toPrinceton in 1975, became interestedin models for neural networks andbrain function about 1979-80. It wasnatural for him to realize that com-plex, interconnected systems of sim-ple units could have the "ruggedlandscape," multistable properties ofspin glasses. Using, very ingeniously,an ad hoc and apparently unrealisticassumption of symmetric couplingbetween neurons, John got the follow-ing results:> For a given set of coupling synap-ses (interactions) JtJ between neurons(spins) i and j , the conventionalMcCulloch-Pitts model of neuronalinteractions maps onto a "greedy"algorithm for finding the local groundstate of a corresponding spin glass.("Greedy" is the computer scientists'
Philip Anderson is a condensed mattertheorist whose work has also hadimpact on field theory, astrophysics,computer science and biology. He isJoseph Henry Professor of Physics atPrinceton University.
self-evident jargon term for jumpingdirectly to the lowest local energy foreach spin.)> Modifying the couplings, or choos-ing the Jtj, in such a way as to form theso-called Hebb synapses makes theneural network into a model for"content addressable" memory: amemory like our own, which canreproduce full detail from fragmen-tary information. A system of Nneurons connected by N(N— l)/2symmetric synapses can rememberabout N/6 N-bit messages in the formof locally stable "spin" (that is, neuronfiring rate) configurations. (Johnmade this conjecture about the capac-ity of a neural network on the basis ofsimulations, but it later turned out toagree with exact analytic theory.)Thus, in exchange for a capacityreduction of a factor of 3 relative to theinformation-theoretic maximum, onegets the content-addressable feature.> Finally, several other brain func-tions, such as pattern recognition,could be modeled with the spin glasstype of neural network.
Many of you may be aware of thegigantic growth of neural networkscience in recent years. In 1979,however, when I tried to whip upinterest in John's ideas among com-puter scientists at Bell Labs, therewas little response; and he, AlanGelperin and John Connor got nearlyequally short shrift among neurosci-entists. Nowadays the neuroscien-tists and computer scientists like topoint to prior claims for each compo-nent of John's achievements. I canhardly believe, however, that suchfurther developments in neural net-works as the revival of the perceptronwould have occurred except as aresponse to John's beautiful demon-strations that, after all, one suchsystem—the Hopfield neural net-work—does work and has a rigorous,mathematically respectable basis. Inparticular, John's work has generateda very healthy trend toward rigorousmathematical demonstrations of lim-its on capacity, accuracy and so on inneural networks and perceptron-likemodels, using the statistical mechan-ics methods provided by Toulouse,Haim Somopolinsky, Miguel Vira-
soro, John Hertz, Richard Palmer(who was John's associate at Prince-ton in 1975-78) and many others. Itturns out that statistical mechanicscan be applied to realistic, asymmet-ric networks as well, and that there isno real difference between the capa-bilities of symmetric and asymmetricnetworks.
I promised I would close my serieson spin glasses with this column, so itmust be descriptive, not detailed. ButI must also mention how the "ruggedlandscape" of spin glasses relates totheories of biological evolution. In1981 I visited John Hopfield at Cal-tech and helped with the course on"physics of information" that he,Richard Feynman and Carver Meadwere giving. Stimulated by John'swork on neural nets, I came back toPrinceton with the realization that Icould put my own rugged-landscapeideas into a theory of prebiotic evolu-tion that Daniel Stein (now at theUniversity of Arizona) and I werealready working on. The genome ofan organism can be thought of as a setof Ising spins—two for each base in theDNA because there are four types ofbases and the Ising spin has twopossible values. The fitness, or repro-ductive capacity, of the genome can bemodeled by a frustrated, quenchedrandom function of this list of spins,and the simplest random functionthat satisfies the requisite plus-minussymmetry is a spin glass Hamiltonian
•function. (The plus-minus symmetryis imposed by the complementarity ofbase pairing.) With a senior thesisstudent, Dan Rokhsar, Stein and Imade a simple model of primitiveevolutionary processes using this idea.
Related ideas, but without the sta-tistical mechanics insights, had al-ready occurred to Gerard Weisbuch atthe Ecole Normale in Paris and to StuKauffman at the University of Penn-sylvania. Nonetheless the extra un-derstanding those insights providehas encouraged us, and especially Stu,to go on and attack all kinds ofevolutionary—and other—problemswith spin-glass-like random, rugged-landscape models. This approach hasbecome an important part of theprogram at the Santa Fe Institute, of
1990 Americon Insrirure of Physics PHYSICS TODAY MARCH 1990
SPECTALEEPREVERSE VIEW LEED / AUGEROMICRON are now the world leader in Reverse View LEED / Auger systemswith over 250 delivered worldwide. Our continuing policy of development
has enabled us to provide new features and improved performance.
• LEED controller
The 1990 systems features:• A new high intensity LaB6with identical run-up timesto those of standard filament types.
i New, low divergence miniature electron source.• Low cost Auger datasystem now available, based onIBM PC or compatible., .
• Complete Auger electronics package.i Channelplate version offered for low primary beamcurrent operation.
. Auger control package
INSTRUMENTS FOR SURFACE SCIENCEWEST GERMANY:OMICRON VAKUUMPHYSIK GMBHIdsteiner Str. 78D-6204 Taunusstein 4Tel. 0 61 28 /7 30 44Fax, 0 61 28 /7 35 96Tlx. 4 182 769
USA:OMICRON ASSOCIATES1738N Highland RoadSuite 208Pittsburgh, PA 15241Tel. (412)831-2262Fax. (412)831-9828
UK:CABURN UHV1 Castle Ditch LaneLEWESEast Sussex BN 7, UK.Tel. 0273480081Fax. 0273480545Tlx. 877412
JAPAN:ASTECH CORPORATIONTakadanobaba 21 Bldg No 39-7Takadanobaba 4-ChomeShinjuku-ku, Tokyo 160Tel. 03-366-0811Fax. 03-365-5577
ISEE US AT MRS, SAN FRANCISCO, BOOTH 7021Circle number 8 on Reader Service
REFERENCE FRAMEwhich Stu and I are members. Unfor-tunately I cannot discuss here themany other ramifications of this wayof thinking. From my point of view,its attractiveness lies in that it allowsus to explain simultaneously the con-tradictory aspects of variety and sta-bility of certain special forms andpatterns. Life exhibits these contra-dictory aspects: Among the countlesspossible mutations, many lead to sta-ble species. In the language of spinglass theory, there are many "basinsof attraction," but each is stable (or atleast metastable). It is also clear, asGerard Weisbuch first pointed out,that evolution in such a landscapewill exhibit "punctuated equilibri-um," or sudden changes from onedeep maximum of fitness to another, afeature that has been emphasizedrecently as characteristic of much ofevolution.
I realize that I never returned to"real" spin glasses, even though itwas studies of the low-temperatureproperties of those dilute magneticalloys that led to the theoretical ideasI have been discussing. There is areason: In spite of much beautifulexperimental, computational andtheoretical work, a complete and con-sistent understanding of those mate-rials is not yet at hand. HeleneBouchiat and Pierre Monod, amongothers such as Laurent Levy, havebeautifully demonstrated that realspin glasses have divergent nonlinearmagnetic susceptibility at TSG, verify-ing that there is a real phase transi-tion, albeit one without a visiblespecific heat singularity. Spectacularsimulations carried out on specialpurpose machines by Peter Young(University of California, Santa Cruz)and Andrew Ogielski (AT&T BellLabs) have also verified the existenceof a phase transition in three dimen-sions. Daniel Fisher (Princeton) andDavid Huse (Bell Labs), among others,speculate, however, that real spinglasses are really not ultrametric orreplica-symmetry breaking. The the-ory is still under development. Someof it was explored in the December1988 PHYSICS TODAY special issue ondisordered solids.
For further information on randomlandscapes and evolution, Stu Kauff-man's forthcoming book Origins ofOrder: Self-Organization and Selec-tion in Evolution (Oxford U. P., NewYork) is perhaps the best source.John Hopfield has written (with Da-vid W. Tank) a Scientific Americanarticle on his neural network ideas(December 1987, page 104). An arti-cle on neural networks by HaimSompolinsky appeared in the Decem-ber 1988 PHYSICS TODAY (page 70). •
VCH
see these &other finephysicsbooks at
APSBooth #418
For a completePhysics Brochuredetailing all VCH
physic's books, pleasewrite to:VCH
220 E 23 ST Suite 909NY, NY 10010-4606
Customer ServiceOrders / Inquiries: VCH
303 NW 12th AveDeerfield Beach,FL 33442
In a hurry? Call.A MC> ann Ann aanA
In Canada: 305 428 5566
Outside of North America,Please Contact: VCH
PO Box 101161D-6940Weinheim (FRG)
ENCYCLOPEDIA OF PHYSICSSecond, Revised EditionEditors: Rita G.Lerner & George L.TriggReviews from the First Edition...
The reference value of this volume has been enhanced by the liberal use
of cross-references and the Inclusion of a good Index. Almost all the
entries contain 'see also'references...Index Is both comprehensive and
accurate.lt Is recommended as the most recent authoratative reference
source on physics for libraries specializing In sciences and all larger
public libraries." -American Library Association
"I found that almost without exception the contributor for a particular
entry was among the top five that I would have listed as best qualified to
write on the topic. This Is clearly the definitive encyclopedia of physics
currently available...The format, two columns per large page, is
convenient...attractive and easy to use." -American Scientist
This will rank as one of the best one-volume physics reference books for
a long time." -CHOICE
Thoroughly updated second edition of one of the most successful referenceworks of physics ever published in one volume. Over 500 articles by authoritiesin physics, including several Nobel prize winners. Covers all aspects of physicsfrom astrophysics, biophysics and elementary part'icicles to xerography.Expected September, 1990 ca. 1408pp. Cloth 73752-3 List: $150.00
announcing...ENCYCLOPEDIA OF APPLIED PHYSICSEditor: George L.TriggThis comprehensive encyclopedia will consist of 18 volumes, arrangedalphabetically covering all aspects of Applied Physics. The approximately10,000 pages spanning the entire scope of physics will be written by interna-tionally recognized experts. First volume is expected January, 1991.For more information, please write to us at the New York address below.
A RANDOM WALK THROUGH FRACTALDIMENSIONSBrian H.KayeLavishly and beautifully illustrated introduction to the fascinating world of fractalgeometry is intended as a "first reader" for those who want to apply fractalsystems to their work. Contains numerous graphic examples and suggestionsfor experiments.1989 421pp. Cloth 73496-6 $85.00 Paper 73888-0 $39.50
P H Y S I C A STATUS SOLIDI (A & B)Editor-in-Chief: E. GutschePHYSICA STATUS SOLIDI (A&B) contains important results from both basicand applied research as well as the latest experimental techniques. Two issuesconstitute one volume of roughly 900 pages. Series A (applied research) coversmaterials science & applications; solid -state diagnostics & other methodologi-cal aspects; solid-state phase transitions; and device-related phenomenaSeries B (basic research) deals with such topics as electronic statesand tran-sitions; lattice dynamics; disordered, lower-dimensional and organic systems-surfaces; and high excitation phenomena. Journal is indexed in CurrentContents/ Physical, Chemical and Earth Sciences. Akademie-VerlagPublished Monthly (A) 0031-8965 $850.00 (B) 0370-1972 $850 00
Circle number 9 on Reader Service CardPHYSICS TODAY MARCH 1990 1 1