June e@2, 1974 PAGE 1

SUMEX#AIM PROSPECTUS

INTRODUCTION AND OBJECTIVES

In partmershio with and with fimancial support from the
Biotechnology Kesources Branch (BRB) of the Division of Research
Resources, National Institutes of Health (NIH), Stanford
Untversity is develoning and operating a National Shared Computing
Resource to explore advanced applications of computer science in
health research, Tre SUMEX (Stanford University Medical
Experimental Camputer) facilitys under Professor Joshua Lederberg
as Principal Investigator, {s national in scope tn that a major
part of its computing capacity will be made available to
authorized research groups throughout the country by means of a
communications metwork, There are two main objectives of the
facilityrt 1) the specific encouragement of apelications of
artificial intelligence im medicine (AIM) and 2) the managerial,
administratives amd technical demonstration of a mationa) shared
techmoloaical resource for health research,

The emeraence of more economical technologies for data
communications allows a lihberattan from qeogranmhically organized
facilities in favor of the concentration of fumctionmally
specialized capabilities at different sites, In addittom ta the
economic advantages of resource sharing, a closer interaction
between diverse research efforts is expected to promote a more
systematic exchange of research products and ideas, This is
particularly true in comouter science where technmical complexities
have tended to encourage the development of relatively isolated
groups, each pursuing a line of pesearch and program development
and making limited use of the working programs availahie from
others, Tre SUMEX@AIM project seeks to lower these barriers in
the specific area of artificial intelligence appiied to health
research, Indeed, nultilateral communitys=bulilding pather than
unilateral service is the project!s essemtial mandate,
"ARTIFICTAL INTELLIGENCE” PESEARCH

The term "artificial imtelligqemce" (AI) ts applied ta
research efforts aimed at studying and mechanizing information
Processing tasks that have generally been considered to require
human intelligence, The cupremt emohasis in the field is on
efficient acauisition and utilization of matertal knowledge, and
Pepresentation of comeeptual abstractions in problem salving
processes, AI systems are characterized by comolex information
processes that arer to a large extent, momemumerics, OrGer graph
searching and symbolic patterm amalysis, They involve procedures
whose execution is comtrolled by different types and forms of
knowledge about a aiven task domain, such as models, and fragments
of "advice" im the form of systems of constraints or heuristic
pules, Uniike canventional algorithms commonly based on a well~
tatlored method for a aivem task, AI procedures typically use a
multdolicity of methods im a highly conditional mamner = depending
om the specific data in the task and om a variety of sources of
relevant information,

For example, an AI system for data interpretation would
assist a user at a higher conceptual level tham the application of
a numerical algorithm fore eurve fitting by suggesting and
evaluating analytical expressions to match the data, Thus it may
include decidirg what model best aeseribes the data source and
what actions to take as a result, Such a model would be based an
the data itself and on other pieces of tmnformation which may be
relevant to constraining a solution, Inherent im these precesses
are a store of general symbolic knowledge about the problem domain
describing what ig fact, what 18 reasonable and unreasonable, and
what solutions have worked in the past, Using an ability to
commumicate effectively with the problem data source and a human
uSer, the programs might examine the knowledge base to constpuct
and test plausible explamations of a particular set of data and to
project and decide among subsequent alternative courses of action,

This type of “tntelligent" assistance by computer program
is perhaps best illustrated by a number of examples taken from on»
goima research efforts identified as initial users of the
facility, The DENDRAL project at Stanford !s aimed at assisting
the biochemist im interpreting molecular structures from mass
Spectral and other chemical information. In cases where the
characteristic spectrum of a compound is not catalogued im a
library, these programs carry out the pather laborious processes a
chemist must go through to interoret the spectrum from "Fterst
principles," By symbolically generating "reasomable" candidate
structures from himts within the spectrum amd a kmowledge of
organic chemistry and mass spectrometry, the program infers the
unknown structure to be the one which best explains the observed
spectrum, There 18 no direct algorithmic path avalilanlie to
Getermine such a molecular structure from the spectral data = anly
the inferential process of hypothesis generation and testing
within the comain of reasonable solutions defined by a knowledae
of orgamic and physical chemistry,

This process, aS implemented im the computer, is a
simplified example of the cycle of imductive hypothesis =
deductive verification that 48 oftem taught as a model of the
scefentific method, (Whether this ts a faithful description of
comtemporary science is arguable; and how it may be implemented in
the human brain fs unmkmown, In any case, these are useful leads
pather than absolute preconditions for the pragmatic improvement
of mechamized intelligence for more efficient probiem solving,)
The elaboration of these approaches as deeply as we cam with
existing hardware and software technologies is the most promising
approach to enhancing the anplication of the computer to the
vaguely structured problems that dominate our task domains,

A project related to DENDRAL, carried out in collaboration
with the University of Califormia at Sam Diego, seeks to inter the
structures of complicated proteins from x=ray crystallographic
data through a similar paradigm of hypathesizing reasonable
Structures and testing thefr ability to explain the observed data,

Uther projects at Stanford and Rutgers University seek to
assist experimentally tn déaqnosing disease and suggestina
treatment, In these cases the input data are clinical symatoms
and physical and biochemical measurements and the knowledge damain
is the physiology and pathology of various organ systems as wel]
as the effects and interactions of courses of treatment, Current
applications of this work are im the areas of infectious diseases
and diseases of the eye, Long term applications of these tynes of
computer programs might be to consolidate and reconcile the
kmowledge from a diverse group of experts or to enable more
effective treatment of disease im locations that lack access to
specialized exrertise,

Cther exammles, im areas of psychology, are aimed at
bulidimng and testing complex models of human cognitive and
affective processes, Programs are being designed at Stanford to
simulate paranoid or other behavior patterns in rasponse to
natural language discourse, Embedded im such programs are a
general symbolic model for the behavior type and am ability for
matural language communication, Im response to a kumam user, the
program seeks to understand input discourse im terms of the
behavioral node! and to oroduce aprropriate Enagiish language
sentences im response, Applications of this work may be to
improve our understanding of particular behavior patterns by
systematizing the characteristics of models which emulate them or
to assist im training medical students im psychiatry,
A system is heing developed at Rutgers which takes a8 an
imout a social entsode, {,e,5 an account of a sequence of actions
involving the interactions of several persons within some social
context, and it generates an interpretation of the episode in
terms of imtemtions and reasons that might have motivated the
persons im the episode to perform their actions, This system is
based om a model~comsisting of a body of rules of belief (about
scecific people, actions, ard motivations) ard a strateav of
interpretation, Applications of the system may be in improving
commumication processes such as the psychiatric interview,

These examples are given for the purpose of comcrete
filustration of the types of problem areas we seek to explore,
They are mot intended to bound the domain but rather to stimylate
new ideas along the lines of "Intelligent" programs with medical
applications, These terms are mot precisely defined at this time
and ovr objectives miaht better he phrased {nm terms of "advanced
computer science comcerned with mechanized theory formation and
problem solving {m medical research and practice," we will
emphasize the past achievements of AIl=oriented and Al=laheled
Pesearch in providing facilities for further advances such as
symbolic knowledge representation and manipulations comcept
formations problem solvings learnings and matural human
communication (e,g, language, speech, vision, etc,),

MANAGEMENT AND USER QUALIFICATIONS

The SUMEX*AIM facility is community*oriented ana fits
OrFganizatienal structure jis being established to emphasize user
support and interactions among user groups, The following
summarizes the overal! structure as ft is tentatively comstituted,
Additional detail will be srovided as appropriate, giving specific
information about how prospective users may gqaim access to the
facility,

The user commyrpity is divided for administrative purnoses
imto two groves: 1) these at the Stamford Medical Schoo! (loca)
to the facility) and 2) those elsewhere {mn the country and at
Stamford, The facility resources (computing capacity and
manpower) will be allocated {in equal portions to these two groups,
Stanford Medical Sehoo!l users will be qualified for access to the
facitity by Dre, Lederberg in Kis capacity as Principal
Investigator for the SUMEX areant, The national user group will
gaim access to and be represented in the aesign, development, and
allocation of the facility pesources through am advisary group for
@ mational Program in Artifieial Intelligence in Medtcine (AIM),
The AIM Advisory Group will comsist of members at large of the ATI
and medical communities, facility users, amd the Principal
Investigator of SUMEX as am ex=officio member, A representative
of the NIH*BRS will serve as Executive Secretary, It will advise
the AIM Executive Committee, whose responsibility ts to aive
overall direction to the matiomnal AIM program, The AIM Executive
Committee comsists of the Principal Imvestigator of the SUMEX
Projeets the Principal Investigator of a series of workshops on AI
applications im medicine, a representative of the NIH=ARA, and a
represemtative of the AIM Advisory Group,

Besides its charae to review potential uses of the SUMEX=
AIM facilitys the AIM Advisory Group wi!) advise the AIM Executive
Committee om the allocation of funds needed to assist in
imterfacing Pecommended new users, The SUMEX=AIM computing
resource will be made available to qualified users without any
charger which cf course entails a careful review of the merits amd
eriorities of proposed applications, At the discretian of the
advisory group, attendant commynication and transportation costs
to allow specific users to gain access to the facility may be
covered as well,

Qualifications for new users amd the details of NPAIM
assistance to users will be more specific at a later time, In
general terms, however, potential users will be judged on the
basis oft?

i.e The sctentific interest and merit of the pronmosed
research,

2, The relevance of the work to the AI approach of SUMEX=AIM
as may be tnmdicated im part by the meed for the specialized
SUMEX facilities as opposed to other computing alternatives,

3, The prospective contributions and role of a user in the
COMMUPMATYs @,9er developing and sharing new systems or
applications proaqrams, sharing use of special hardware, etc,

4, The user's capability and dntentions of operating tn a
communmityseffective style for mutual advantage, Besides the
programming innovations that some users may be able to
contribute, atl are expected to furnish expert knowledge and
advice about the existing art im the fields relevant to their
specia) interests,

Tee averal!l objective will be to promote AI applitecations
of high acientific merit among an extensive group of competent
users, comsistent with maintaining a pesmomsive and productive
comouting environment, The imitially approved loading is
estimated to he about 30% of the total SUMEX@AIM capacity, leaving
a substantial capacity for mew research groups to enter at this
time,
TECHNICAL CAPABILITY

The comouter facility, consisting of a Digital Equtoment
Corporation mode) KI#10 central processor operating under the
TENEX time=sharing monitors, {8 Scheduled to be available tn a
limited fashion im Jume 1974, and to be fully operational in
September, The system will have initially 197K words (36 bit) of
hign speed memory}; 1,5M words of Swapping stcrage: 40M words of
disk storagey 2 9=track, 800 boi industry=compatible tape umits; 1
duel DEC tape unit: a lime printers and communications network
interfaces providing user terminal access, At the present time
the cholee of a communications network approach is being
fimalized, However, specialized users will im any case find that
the commumications costs are smal} comeoared to the value of access
to the system, This hardware complement may be expanded in the
future based om available funding and Justified user community
needs,

Software support will evolve from the basic system as
dietated by user research goals and requirements, Initially
available programs will include a variety of TENEX user, utility,
and text editor programs, Major user languages will inctude
INTERLISP, SNOBOLs SAIL» FORTRANW10, BLISS=10- BASIC, and MACRO#

De, Joshua Lederberg

SUMEX=AIM Computer Project
Stamford University Medical Center
Stanford, California 94305

(415) 49765801