A Summary Treatise on Sentient Machine I:

The Cognitive Abstraction Inference Induction Machine.



Written by Jack A. Shulman,
Director, Sentient Machine Laboratory.
© Copyright 1996 Jack A. Shulman, All Rights Reserved.

THE MORPHOLOGY: HOW WE THINK.

Imagine, a machine which can think like a human. A Semiotic Machine. A Sentient Machine. Humankind has fantasized about the creation of a mechanical man since the ancient Greeks, since DaVinci, since the Indian Educational Institutions of the Hindus Valley 5000 years ago.

The notion of such an "electronic brain" ranges from the imprecise ("Electronic Brain predicts Eisenhauer in Landslide victory over Adlai Stevenson - beats human mathematicians" - screams a New York Times cover article edited by Alfred O. Sulzburger during the late 1950's) to the precisely disingenuous ("Sleeze indicted by New York State Attorney General for driving expensive sport cars and claiming he can build a computer that can think like a human being to investors! Brokerage being operated out of a back room, says Deputy AG" - screams a New York Daily News article written by one David Glick in 1986).

Major computer companies spend fortunes on the concept, yet they can't quite seem to find it within themselves to understand what it really takes to build such a device ("IBM Deep Blue computer to challenge World Chess Champion - its Man vs. Machine").

Synthetically, a machine which duplicates the operation of the brain in every aspect may be an exercise in overkill on the semiconductor plane of existence. However, it can be clearly demonstrated that to understand the Human Mind, one must first understand the semiotic processes thereof, and to do that requires a correlative mapping of the brain's physiologic structure to the human body and mind's operation. Accordingly, the underlying principles of human thought, implemented in semiconductors and software, may be a more effective approach to implementation of the Semiotic structures of the human mind, than relying solely on mimicking the grandiose effects of nature by building an "Android Brain" with every detail of the Human Brain recreated in all of its fine detail and function.

The Cerebral Function and the Medullar Function - inter-cooperating at the Hypocamul Index, directed by the limbic and neolimbic or brocal operations, produce a pronounced reasoning method/systemology you and I call "thinking".

Thinking is an artificial process. The Mind itself doesn't think without first learning to think, it learns to think during the first four years of life, laying down a half million new connections each minute during the first few hours of life. After all of the semiotic processes which result from this learning are completed, then the mental process, or mentation, as some are given to call it, begins to make sense and thinking - that is, mentation which makes some form of sense, commences.

Sometimes, this takes on the form of listening to the inner mind "talking", sometimes it takes on the form of constructing a list and eliminating all but the solution, or no solution, sometimes it takes on a massive search operation which is directed in a manner which causes the most promising matchings to a situation to take on louder rather than softer tones of the mind. Sometimes it does other "things".

These "things" are among what the CAII-ist calls: "implementation protocols" of thought (IP).

Semiotic processes are the elementary components which are strung together by the mind to produce IPs. The mind adapt many semiotic processes into a single learned method for solving a class of problems when presented to the mind.

In the formulation of semiotic processes (sub-protocols of the IP), the learning processes (or in the case of instinctive behavior - pre-defined semiotic processes) are tested for their relative usefulness in the learning program, and are identified as "abstracts" bearing conceptual or organizational laden information. These are stored in groups of colloidally related neurons, with long term paths defined between them, so that they may be recalled by other structures of the brain.

The semiotic process (or sub-protocols) are themselves composed of networks of weighted instantiations that we call "foundation objects" or the fundamental mechanisms.

THOUGHT: "T=C * A * I(Squared)"

There are only four fundamental mechanisms used by the mind: CAII, each a basic pre-semiotic process. Each has a weight driven chemical encoding associated with it that enables the mind to bind them together into a pre-semiotic CAII structure. Each CAII is combined into a hierarchical recombination by the experience indicators encoded into chemical receptors in key neurons within a particular structure. These enable the mind to "recall into" a semiotic process without anything other than "intent" or "motive" and each semiotic process is availed to implement a high level implementation protocol. The structure used to combine CAII instances is called "mimickings" - the creation of by larger scale CAII processes which "mimic" the encodings stored in the lower scale CAII processes. If the intent exists to "keep thinking" along a particular thought thread, these are expanded again to build a higher level implementation sub-protocol, so-called surround layers evolve, and these also use a mimicking differentiation (decomposing the lower level's encodings into individual CAII processes), accruing additional information as the expansion occurs.

Interesting, some such structures, once bound, percolate for very long periods of time, resulting in an answer protruding into the conscious mind even though the mind may not be aware of a long term ongoing deliberation. It often depends upon whether the mind has a motive for long term deliberation.

Often Subliminality or unconscious semiotic processes take place below the level of the listening or sensory consciousness of the brain, at high wavelengths imperceptible to the outer regions of the brain. These appear chaotic or white noise-like to the analyst of brainwaves. They have an analog connection to digitization which defies radiology. It is theoretically true that individual long term complexes of implementation protocols may take on a life or their own (the multiple personality problem) or even allowing the paradigm of L. Faraday - the "Nuosphere" to persist in a commonness of mind among all or a part of humanity.

The mind recognizes classes of problem or situation using a technological triumph of human reasoning implementation protocols called (general) "Cognition". Cognition is a paramutual process connected to the senses (as input) the memory (for recall) and the primary survival chords (in the Medulla Oblongata and the NeoLimbic Brain). Cognition has a large number of IPs defined for it, and may bind itself into a neuronal mass to reply to any form of stimulus - using encodings in the cognitive process to differentiate - signals sent to pre-defined cognition IPs begin echoing Recognitive Images (RI). Cognition stores these RI's within "abstracts", as defined in the following sections.

A long cycle cognitive process may peak and use 1/4 to 1/2 of the brain's total functional mass during the pre-Cog state.

Cognition elucidates the elements of the class of stimuli via a mechanism called "Abstraction" - that is the pulling apart of familiar components into abstracts, each having a meaning to fundamental alphabetic componentry, as in "The Semiotic Alphabet" of J. Norseen [1] of mind and memory. The brain rapidly cross references the abstracts using an associative agreement-disagreement algorithm-like fundamental mechanism called (general) "Inference".

Inference is used here differently than the classical concept of inference - "what does this mean" asks the mind of its Inference receptors and they respond with possible fuzzy responses (FR's) from recall. This results in the foregoing "expansion hierarchy" - the Inductive or Induction primitive takes all of the FR's and "clones" CAII hierarchies, each bound to a particular significance in the stimulus. This process repeats, and repeats, until a semiotic process evolves and is (a) stored for future use; (b) rendered semi-permanent (needed for survival); (c) rendered permanent (this is called "made grooved"); or (d) discarded; during which time the Mind either capitulates into an automatic stream, concludes a conclusion as part of the inductive process or dismisses the CAII/SP/IP as pointless. In the misbehaving mind, entropy can take place which seriously damages the cloned or grooved renderings of each IP, leading the mind to misconceive survival purposes for a particular decision, or leading to a faulty or dangerous reaction induction, such as self harm or illegal behavior.

At every level of the "expansion hierarchy" that the brain creates, "Cross-referencing Pathway Axons" comparate multiple signals using intermediate neuronal masses to test and then signal "Inductive" motivation - they either carry the solution or a partial solution, or continue to direct expansion until ambiguity or confusion arise, or the solution appears to be zeroed in upon in the observation of assigned inductive neurons. If the solution to a problem requires the process to continue during all, or nearly all hours of mental awareness - the mind will normally wire such an IP to an active thread consciousness (or in unconscious mind).

High level Inference, Abduction, Premise Coordination - etc. - these are all high level implementation protocols which are based upon CAII with various quantified variations derived from experiential learning.

The above simplified explanation of the thinking brain is what occurs in simultaneity within the brain when it is is working on several problems or situations at the same time. It may be navigating a Tennis Court while tracking the movement of the Tennis Ball. It may be thinking about a TV show and engaging in a conversation. Coordinated and discoordinated reasoning paths are laid down in the brain to equate differential or non-differential acts, such as hand-eye coordination - yet they take up a large number of neuronal masses just to perform a simple function - while smaller regionalizations can take upon themselves more sophisticated tasks.

The mind implements many such implementation protocols it creates through the Learning Process of Life. Each is a "cantilever", " multilayer" or "seeming random network" of CAII process "remnants" or "artifacts" stored in memory which the mind recalls, activates and "thinks through" against a given problem or situation, each bearing meaningful concept laden abstractions.

THE MOTIVE: WHY WE THINK.

Its simple. We survive therefor we think. Or rather, vice versa. Those species who don't think are not necessarily equipped to resist the threats to their survival, blindly swimming into the mouth of Moby Dick.

We think as a process of physical living, as it supports the living process. We rarely fully understand the results of our own thoughts, as we have thresholds of listening built into the thinking/listening process which tends to filter out only the more important thoughts.

Our entire personality is made up of thought processing. We simply are the embodiment of our IP's and Recall Store.

Despite the modern social phenoms, we do admire thinking. We demonstrate our appreciation by advocating the arts, music, poetry, drama, theater. We debate among our politicians. We award degrees, diplomas, credentials.

And yet, not one of these things necessarily accurately quantify our thinking abilities. Some of our brightest thinkers devote their entire life to profiting from buying and selling on the international stock exchanges, or playing a single game like Chess their entire lives. Perfectly brilliant people run for public office. Others strive to head up a major corporation.

Why? Is it from power or money motivation?

Or is it the mind perceiving a source of survival assistance. If we were to eliminate nearly all the basic or threatening problems (money, job, procreation) from daily life - it might change the way we think as we might have more room in our lives for entertaining notions that are today displaced with plain and simple surviving.

We think to survive or do we survive because we think? Cogito ergo sum notwithstanding - many things which don't think do exist - perhaps We Think therefore we Survive and that is what "therefor I am" means. For were you and I not to think, doubtless our chance for survival would diminish.

CAN or SHOULD A MACHINE THINK?

The real question here is "will a machine which calculates in a manner which emulates human thinking" actually be a machine which thinks like a human?

In reality, it will be a machine which in many millions of and billions of machine instructions executed each second, implements software programs which themselves implement that which appears to operate like a human thinks. Is it thinking? The answer is: if the stimulus is the same - and the response the same - the program works, but it does not really think in the same way humans do: it thinks "like" a human, not exactly the same as a human.

THE EVOLUTION: HOW CAN A MACHINE BE MADE TO "THINK LIKE A HUMAN"?

Well, to begin with, one must design the software capable of doing so - software which implements Cognition, Abstraction, Induction and Inference fundamental mechanisms, software capable of building pre-semiotic processes out of expansion hierarchies of CAII and, further, then implementing various reasoning methods which derive from same.

The disposition problem - how do we define a personality for such a machine - one for the human interface experts who today limit themselves to conceiving of "windows" and "prompting" and "menus" and "hypertext". What are the Sentient Machine equivalents:

1. Listening and Watching Apertures.
2. Pre-Prepared Voice Synthesis.
3. Command Replies.
4. Human-friendly Audio Apparatus.
5. Display panel for providing visual figures.
6. Self Observatioin: Software and Hardware Sensors for examining what the machine is commanded to do, exertion of Inhibition, etc..
7. Direct contact human interfaces.
etc...

Then, one must cast these software programs into a machine capable of executing them extremely rapidly, because hundreds of thousands of these "Software IPs" will have to take place in a single second, and many thousands side by side, to produce a solution to a problem in a similar fashion to the human mind or to learn a method or source of information, in a reasonable enough period of time. This will require a machine of multiple processors in a number of different geometric relationships, each processor of which is capable of several billion machine instructions per second - because each fundamental CAII mechanism itself may take hundreds of thousands of machine instructions - and a single implementation protocol requires hundreds of billions of linear and parallel instruction executions in a given per second. Software must be capable of linear, non-linear and hyperbolic problem solving, which places enormous back propagation loads on a machine which is insufficiently assisted by algorithmic advances, such as neuroconnectionist hybridization of reconfigurable hetero and homeo morphic processor arrays. An example of a more advanced mathematical approach to topologies, Transfigural Math (Matran Logic) evinces the possibility that for specialized higher application processes, even a special processing unit, such as a Matran Logic Arithmetic Processing Unit, could be used to shorten many of the traversals required to solved certain classes of problems.

WILL SUCH A MACHINE BE SENTIENT OR SELF AWARE?

The term sentience is often thought of as "self aware". I am not sure that any machine has the capability of true self-awareness. The Sentient Machine will be relegated to trying to achieve "Synthetic Self Awareness". However, the software within that machine may very well be capable of emulating self awareness during its effort to "think like a human".

Cogent? Probably. Sentient? That remains to be seen.

However, do not be surprised if machines which emulate sentient beings, at least in appearance, begin to appear at the turn of the Century.

ONE DESIGN.

The Sentient Machine design has followed many different project requirements in the past.

For example, the following is extracted from one of my design exercises for a client, a machine called "the MPPC Machine" designed for a Government project: "Nightwing", released by permission of the commercial client, taken from a resulting study by a researcher, who analyzed one of my past machine designs and came to his or her own conclusions about it.

------------------------------------------------------------------------

AN EXERPT TAKEN FROM A RESEARCH REVIEW OF A PAST PROJECT

------------------------------------------------------------------------


Phase Two: Design Examples of Shulman Advanced Sentient Emulation

The intended design program takes function in two formats:  a) A centralized Massively 
Parallel Processing Complex (herein called "MPPC") and, in deference, b) a Massively 
Parallel Graphics Workstation (herein called "MPGW") which utilized parallel processing 
within its architecture.  Both would  be designed to incorporate massively parallel scaling 
as in TABLE 1 below at levels "64k to 256 m".  Each system (the MPPC and the MPGW) 
is capable of independent operation, and a subsystem (the Hyper System Gateway, or 
"HSG" herein) would be designed to allow the assembly of a so-called "HyperSystem".   
A Hypersystem is one which would consist of a number of MPPCs (used as compute 
servers) and a number of MPGWs (used as workstations for various points of attack 
in an application.  These systems are capable of being assembled together in varying 
number, depending upon application, interconnected by an HSG of sufficient capability 
to allow free and independent cooperation among tasks in and among the various 
subsystems of the MPPCs and MPGWs.  The HSG are designed to enable use of 
various long distance communications facilities to host a community of MPPCs and 
MPGWs which were locationally diverse.

NOTE: The massively parallel Graphics Workstation is being developed for the 
Department of the Army under Research Topic A93-356. 

The advanced attack/defense and targeting training system for the MPPC is being
developed under a different topic and is classified.

The Centralized Massively Parallel Processing (CMPR).  The Centralized Massively 
Parallel Processing Complex was developed for the Army under Research Topic A93-191.  
The writer feels strongly that there should be open interoperability between these two 
styles of system, as Shulman details in his pre-design stage study, and hence plans to 
develop both whether or not funding is obtained.  The third component, the HSG, designed 
to couple and communicate between and among local or diverse clusters of such systems, 
may or may not be applicable to other final Nightwing Topic's of interest of the Department.

All of the Nightwing designs were licensed from Mr. Shulman and use his multiple to single to 
multiple buffered/cached memory scheme for memory sharing of data between dissimilar 
components.

The parallel computing products include certain mechanisms for which original design 
examples are available from the Company, in order to show how such were perfected. 

Two classification pages are provided herein to assist the DOA in determining what 
Mr. Shulman's private objectives are for parallel processing in the future beyond Nightwing.

It is the Company's stated intention to support Mr. Shulman in seeking these objectives.  
For reasons of confidentiality, only the exhibit pages are provided hereafter.  Accordingly
the Company has issued this report for the purposes of technology transfer review.

The following is a performance table exploring human intelligence capability in 
comparison to machine intelligence, devised by Mr. Shulman.


TABLE I
KEY:
K = 1024
M = 1024 x 1024
SM/MPU = Single or Multithread Microprocessor

NUMBER OFCONNECTIONSM/MPUFACTOR OF
CPU NODESCOUNTSPEEDHUMAN
(VIRTUAL) (VIRTUAL)(REAL)INTELLIGENCE
1 K 3584150 MIPS1 %
16 K57388230 MIPS2 %
64 K229529370 MIPS5 %
256 K918120450 MIPS7.5%
1 M3672488450 MIPS10 %
64 M~14 M880 MIPS90 %
256 M ~58 M880 MIPS99.99 %


NOTE: The speed given refers to the performance of a single node, which is 
multiplied by the number of CPU Nodes in the first column (hence - the CAII Chart at 
256 Million shows that many 880-MIP "VIRTUAL" CPU NODES would provide a machine with 
99.99% of ordinary human reasoning ability at time factor 2 -- twice Human Speed).

The ratio of virtual to real CPU nodes has been excluded from the above, 
except that the number of real is considerably less than virtual in number, by a 
factor of as much as 1000 - due to virtualization intended to lower cost.  Virtual 
CPU Nodes derives from an architectural principle developed by Shulman.  In addition 
to the raw parallelism, the system defined during the above reckoning of human 
intelligence performance, bases itself upon the use of an advanced memory device 
for interaction with the MPU/Memory in the specialized parallel architecture.  
A specialized coherency technique insures that all MPUs are in the same and coherent 
state at all times during operation, and that incoherence in pipelines, fetching 
and translating states of processors are protected from the main execution state of 
the machine.

In addition, the rendering of architectural relationship of nodes to each other, 
to networking, to memory, memory sharing and scheduling, prediction, and data 
transfer paths are all critical to yield the above responses.  These architectural 
principals, and others used in this unique device are revolutionary in their
approach to minimum overhead, minimum distance networking.

The Quadrasphere, used in a Shulman machine reviewed 5 years ago has 
been considerably expanded in scope, to include multidimensional attributes, 
unilateral memory sharing in virtual to real memory translation, and advanced, 
wait stateless prediction of memory and CPU assignment and usage.

His "Synthetic Intelligence" model: Cognitive, Abstraction, Inference and Induction 
has been improved for application to a native computation and parallel
software engine to be embedded within the workstation and complex, so as to 
result in the following state diagram within the machine structure, which 
structure is circular and heteromorphic (multiply recursive and reentrant 
instances), and which returns to its top after completion of the bottom most state.  
Individual states include internal branching and looping, and transfer to another 
basic CAII block, on a sub component by sub component basis.

The interconnections between the states were implemented in state saving 
advanced message passing channels assisted by advanced hardware designed
to provide a true machine architecture, with RISC and CISC microprocessors
used to implement particles requiring variable programming.

The learning modes and active reasoning modes of each state are implemented 
in a pure fusion of machine code and software, proprietary to the Company and 
will remain converse from each other.

The passive and active intuitional elements are implemented as rule based 
and neural network "sub-engines".  A Sub-Engine is a software and data object 
which has the ability of using data to process instructions within what Mr. Shulman 
calls the "Reasoning Sphere".  Each consists of fragments of executable code 
and codified data which are commonly defined within a given class of Sub-Engine.  
The classes of Sub-Engine determine which interoperate with each other in 
different instances of processing nodes connected by the message passing 
mechanisms and software protocol of the system.


TABLE 2


STATE	     NAME                    PURPOSE          DECISION
-----      ------------------      ---------------  ------------	

I	COGNITION STATES

ACQUISITION	Leading Edge State (Distinct)

Significant Implementation Protocols

II         DIFFERENTIATION         Transition       Identity	
III        RECREATION              Capture          Identify	
IV         COMPARISON              ReCapture        Confirm

...

V        ABSTRACTION STATES        Recognition's     Trailing Edge	

Significant Implementation Protocols

VI         TRANSFORM               Transition       Other Identity	
VII        CORRELATION             Capture          Other Identified	
VIII       ATTACHMENT              ReCapture        Conform
	

IX	INFERENCE STATES           The Associative Pre-Semiotic States

Significant Implementation Protocols

X          DISCERNMENT             Leading Edge    Signified	
XI         ASSOCIATION             Transition      Gather Identities	
XII        INVESTIGATION           Capture         Gathered Identity	
XIII       DEDUCTION               ReCapture       Reference	

XIV	INDUCTION STATES           The Connective Post-Semiotic States

Significant Implementation Protocols

XV        COLLECTION              Leading Edge     Possible	
XVI       ENUMERATION             Transition       Order	
XVII      PRIORITIZATION          Capture          Priority	
XVIII     DELIVERY                ReCapture        Induction	

IIX 	EXPANSION (TRANSITIONAL) or TRAILING EDGE (DISTINCT)


Most of the above principles would find their way into advanced 
software Mr. Shulman demonstrated will provide the synthesis of 
certain parts of human reasoning intended for "embedding" within 
the machine, so as to accelerate their use by application software. 

In the Shulman architectural concept a premiere discussion takes
place about semiotic processes.  The fundamental theorem is that
implementation protocols and semiotic processes are two sides of
the same coin, as in the following generic definition of a
unary implementation protocol leading to a "well defined" and
"wired" memory (a concept laden abstraction stored for retrieval).

Process ->  Boundary -> Pre-Semiotic -> 
   Semiotic -> Post Semiotic -> Boundary -> 
      Container -> Memory

These relationships, implemented in the active machine correlate
higher functions like abduction and elimination, explaining the
mental process as "macro implementation protocols" of the conscious
mind which operate like a "language interpreter" - but with some
variance from male to female brain structures - hence the need for
gender based decision making capability in some software.

According to the machine model - the mind is tri layered and should
be implemented so in the human reasoning emulator (HRE), thus:

{Basic Function}
  Fundamental Primitives
  Implementation Protocols
  Defining Identities
  Concept Laden Abstractions
  Safety Processes

{Mid Brain Function}
  Talker / Listener
  Identity Refiner
  Rules and Purposes Management
  Language Interpreters
     -Algorithm Definers
     -Perceptrons

{High Brain Function}
  Application Interface
  Physical Interface
  Control Interface

The above is an approximation of the "culled component list" for
the internal SI process complex. Each "expansion" represents one 
more component, yet uniquely in Shulman we find an understanding 
that the mind's pathways are "disordered" and "weighted", hence 
the HRE must configure itself with an ordering of connections 
routed by the long experience line of the HRE and the pathways 
laid out by pre-programmers.


END OF EXERPT

THE PRESENT DIRECTION OF OUR RESEARCH

From the above it can be seen the direction the Sentient Machine is taking into massive parallelism. Speed as well as structure are crucial. In one proposed design, I combined Neural Connection, Parallel Network and advanced Accelerator Design to produce a Hybrid Neurospheric Structure which was appealing in that it closely resembles the structure of the brain as well as the function, however disassociated in physical appearance it would be if placed in a room next to a picture of same. That is not to say it was identical to the brain, it just had many features reminiscent of such a biological entity.

I feel that Sentient Machine, at least level one, is feasible to implement in under 5-7 years. This strong statement relies on the elimination of the largest single obstacle - the extremely political nature of the struggle to produce it. If it is not, then the Public can be assured that serious commercial interference with research has taken place - as this kind of technology has serious consequences to certain commercial products which it could obsolete.

After that, Cybernetic, Advanced Biotechnological and CyberBioTechnological machinery should emerge at 20 year intervals.

These machines are to be our servants. They are to enhance our abilities - not replace them. It will not be very easy to build such machines in a manner in which they could keep pace with Human Intelligence's rapid evolution.

These machines, carefully designed with human vision and related sense interfaces, must include substantial mechanisms designed to prohibit the possibility of such a machine ever damaging its designers or engaging in acts of aggression or war without an insurmountable master to keep them ultimately in human check.


ABOUT JACK A. SHULMAN. Jack A. Shulman is, quite simply put, one of the fathers of the unique computer technology in our modern era. Jack has provided research and development work for IBM, AT&T, APOLLO/HP, BELL LABS/AT&T Information Systems, PRUDENTIAL, the US GOVERNMENT and dozens of others. Jack is the original developer of the concept of Windowed Graphical User Interfaces and designed the first IBM compatible desktop computer design later adopted by IBM Corporation. He is the father of modern parallel supercomputing and is often consulted with on the subject as a world renowned expert. He is the original designer of, among other, the Proteus, Maisey, Aerosphere, Connection, Nightwing and various other advanced and classified parallel supercomputers. Jack's genius spreads into small, medium and large scale computers and supercomputers from all angles.

Jack is employed as Chief Scientist at the American COmputer Company, and is involved in the international technology development field.

Mr. Shulman is on loan to ACSA where he consults on projects and is director of the Sentient Machine Laboratory.


E-MAIL? Send Email to ACSA [click here] at 72662.133@compuserve.com.