Replacing the Data Processing Computer with Self-Learning Machines
based on the Autosophy Information Theory
Klaus Holtz, Eric Holtz, Diana Kalienky
Autosophy, 602 Mason #305, San Francisco, California 94108, USA
Keywords: Autosophy, Information Theory, Multimedia Archiving, Self-Assembling Data Networks, Self-Repairing
Memories, Content Addressable Memories, Artificial Intelligence.
Abstract: The programmed data processing computer may soon be eclipsed by a next generation of brain-like learning
machines based on the "Autosophy" information theory. This will require a paradigm shift in memory
technology, from random addressable memories to self-organizing failure-proof memories. The computer is
essentially a blind calculating machine that cannot find "meaning" as our own brains obviously can. All that
can be achieved are mere programmed simulations. The problem can be traced to an outdated (Shannon)
information theory, which treats all data as "quantities." A new Autosophy information theory, in contrast,
treats all data as "addresses." The original research explains the functioning of self-assembling natural
structures, such a chemical crystals or living trees. The same principles can also grow self-assembling data
structures that grow like data crystals or data trees in electronic memories without computing or
programming. The resulting brain-like systems would require virtually unlimited capacity, failure-proof
memories. The memories should be self-checking, self-repairing, self-healing, clonable, both random and
content addressable, with low power consumption and very small size for mobile robots. Replacing the
programmed data processing "computer" with brain-like "autosopher" promises a true paradigm shift in
technology, resulting in system architectures with true "learning" and eventually true Artificial Intelligence.
The programmed data processing computer is now
the only known computing paradigm. It is
essentially a blind calculating machine that cannot
find "meaning" like our brain. All data is regarded as
"quantities" according to the Shannon information
theory (Shannon, 1948). An alternative computing
paradigm is now evolving based on the "Autosophy"
information theory (Holtz-Langheld, 1978). All data
is regarded as "addresses". The new brain-like
machines do not require programming or human
supervision of internal functioning. The systems are
essentially error-proof, including self-organizing,
self-healing, and even self-replicating capabilities.
The systems will not produce faulty information
even after severe physical damage.
Autosophy theories and applications
described in many scientific publications available
on the Internet. (Use keyword "autosophy" in a
search engine or go to
A self-learning text database was built in 1988 to
erify the theoretical predictions. Autosophy
Internet television was demonstrated in 2000 (Holtz,
2001). Primitive applications include lossless data
compression in the V.42bis modem standard and
lossless still image compression in the gif and tif
The new Autosophy technology would require a
memory paradigm, which is the main subject of
this paper. Random Addressable Memories, used in
computers, are not suitable. The new mass memory
systems would provide virtually unlimited, cheap
and non-volatile storage capacities, combined with
very low power consumption and small size for use
in mobile robots (Holtz, 2003).
Both the theoretical knowledge and the required
mories are now becoming available for building
brain-like systems (Holtz, 2004). This may evolve
into self-learning robots and eventually into true
Artificial Intelligence. It may also revolutionize all
multimedia communication (Holtz, 2002).
Holtz K. and Holtz E. (2005).
Processing Computer with Self-Learning Machines based on the Autosophy Information Theory.
In Proceedings of the Second International Conference on e-Business and Telecommunication Networks, pages 177-182
DOI: 10.5220/0001412201770182
The programmed data processing computer in Fig. 1
is essentially a blind adding or calculating machine
that cannot find "meaning" as our own brains
obviously can. All input and output data items
(ASCII characters or pixels) are regarded as
quantities according to the Shannon information
theory. Its purpose is to combine raw input data,
using arithmetic, according to a stored program, to
obtain useful output data. The computer cannot learn
and no matter how much data is processed or stored
it will not become more intelligent. All intelligence
is contained in the programming, prepared by human
programmers, where every operation must be
defined by a complex series of instructions. No
matter the speed of the computation or the
complexity of the software, Artificial Intelligence in
a computer will always remain a mere simulation.
Figure 1: Programmed data processing computer
Computers store data in a Random Addressable
Memory (RAM). The address and data content are
both provided by the Central Processing Unit (CPU),
according to the programmed instructions. Data may
enter or exit the computer either via the CPU or
directly via Direct Memory Access (DMA). Both the
program data and the input/output data may be
stored in the same memory device. This may cause
viruses from the communication channels to infest
the computer. A single error in a computer memory
may lead to a total computer malfunction.
The Autosophy theory evolved from research
into self-assembling structures, such as chemical
crystals, living trees, or human societies. These
structures may start from a tiny seed containing
information stored in a DNA helix. The seed will
select specific materials from a random environment
to build itself into a very complex structure without
human design or supervision. In 1974 Klaus Holtz
discovered that the same natural laws and principles
grow self-assembling data structures in memories.
These imaginary data structures grow like data
crystals or data trees in electronic memories, without
programming or outside supervision. The Autosophy
information theory was first disclosed in 1978
(Holtz-Langheld, 1978).
Figure 2: Self-learning brain-like Autosopher
The operations in an "autosopher", shown in Fig. 2,
are based on self-learning networks. The device acts
like a "black box" to absorb all multimedia
information without programming. The sensor
input/output data are regarded as "addresses," which
define or create their own storage locations in the
memory. Computer programming is replaced by
"education" similar to the education of children.
Simple input networks, such as the serial and the
parallel networks, encode the raw sensor data for
further processing in higher level learning networks.
All learning networks may share the same mass
memory device. The memory must be virtually
error-proof for applications in self-learning
intelligent robots.
The serial network, shown in Figure 3, is an example
of true mathematical "learning," according to the
Autosophy information theory. This can be imagined
like the growing of data trees or data crystals where
data input items define or create their own storage
locations in the memory. A unit of knowledge
(engram) is created by new information (GATE),
related to already established knowledge
(POINTER), which may then create a new engram
(ADDRESS) as an extension to that which is already
known (Patent 4,366,551).
Central Processing Unit (CPU)
Arithmetic, Control, Register
Random Addressable Memory (RAM)
Conventional Data Storage Device
Random / Content Addressable CAROM
Self-Repairing Failure-Proof DECAM
Higher Learning
0 Seed
1 0 R
2 1 O
3 2 S
4 3 E
5 2 B
6 5 O
7 6 T
8 2 O
9 8 T
10 1 E
11 10 D
12 10 A
13 12 D
14 13 Y
The learning process can be imagined like the
growing of data trees or data crystals. A stored tree
network is made of separate nodes, where each
ADDRESS represents an "engram" of knowledge.
The library ADDRESS is a mathematical equivalent
to a point in omni dimensional hyperspace. The
content of each library ADDRESS is unique and can
be stored only once. One cannot learn what one
already knows. Because each input data string is
learned only once, the storage requirements would
saturate. The more information already contained in
a database the less additional storage space is
required to store additional information.
There are seven known classes of self-learning
"Omni Dimensional Networks", each providing a
different learning mode. This may include learning
modes that are not available in our own brains.
Some of these learning networks are already
implemented in commercial applications, while
others have been simulated or are known only in
theory. New applications, such as live Internet video
or advanced lossless still image compression, are
now being added at an accelerating rate.
3 E (4) ROSE
2 S (3)
Serial networks store serial data sequences such
as text, sound, or serially scanned images. The
algorithm was invented in 1974 (Patent 4,366,551).
A similar algorithm (LZ-78) was developed later by
Jacob Ziv and Abraham Lempel (Ziv-Lempel,
1978). Most commercial applications use the LZW
(Lempel Ziv Welch) code, a simplified variation
published by Terry Welch (Welch, 1984).
Application examples include the V.42bis data
compression standard in modems and the gif and tif
formats used for lossless still image compression.
Parallel networks store images in a hyperspace
funnel, yielding very high lossless image
compression and fast access to archives. These
networks are especially suitable for archiving and
storage of imaging and video data. Machine vision is
the ultimate application.
Associative networks connect various networks
into a system. They can, for example, connect
questions to answers, text to images, or commands
to execution sequences. A demonstration system was
built in 1978 to verify operations.
0 R (1)
1 O (2)
1 E (10)
2 B (5)
2 O (8)
10 D (11) RED
10 A (12)
5 O (6)
6 T (7) ROBOT
8 T (9) ROOT
12 D (13)
13 Y (14) READY
Pointer Gate Address
MATRIX [ POINTER ] GATE ] (The MATRIX is a working register in the hardware)
Start: Set POINTER = Seed (= 0)
Loop: Move the next input character into the GATE
If End Of Sequence (a SPACE char.) then output the POINTER as a Tip code; Goto Start
Else search the library for a matching MATRIX
If found then move the library ADDRESS where it was found to the POINTER; Goto Loop
Else, if not found, then store the MATRIX into a next empty library ADDRESS;
Move the library ADDRESS where it was stored into the POINTER: Goto Loop
Start: Move the input Tip code into the POINTER
Loop: Use the POINTER as a library ADDRESS to fetch a next MATRIX from the library
Push the GATE into a First-In-Last-Out (FILO) stack
If the POINTER = Seed (= 0) then pull the output data from the FILO stack; Goto Start
Else Goto Loop
Figure 3: An example of the "Serial" self-learning tree network (Patent 4,366,551)
SYSTEMS: Replacing the Data Processing Computer with Self-Learning Machines based on the Autosophy Information
Interrelational networks provide grammatical
language learning that could evolve into talking
databases with speech input and output.
Grammatical speech would be the ultimate method
of communication between humans and machines.
Logical networks yield an advanced form of self-
learning data processing with logical reasoning
capabilities. They may evolve into intelligent robots.
Primary networks provide unstructured access to
archives or databases through deductive reasoning
and automatic indexing.
Hypertree networks promise true brain-like
learning, which is currently being researched. This
ongoing research may add new types of networks
and new learning modes in the future.
The learning algorithms require that the entire
memory must be searched to locate a matching
MATRIX before creating a new network node.
Rapid learning would require a Content Addressable
Memory (CAM) in which all memory locations are
searched in parallel, all in one operation. The
retrieval algorithm, in contrast, would require a
normal Random Addressable Memory (RAM). The
optimal solution is to store each node in two separate
memory devices, one configured as a CAM, the
other as a RAM. This would also provide for
automatic self-repair in failure-proof memories.
When Autosopher robots finally evolve from
mere information access terminals to physical
interactions with human beings, then near-absolute
reliability is essential even in cases of severe
physical damage. A malfunctioning robot may cause
severe damage and injury to human beings. Robots
would require enormous capacity, non-volatile,
Content or Random Addressable memories. The
memory units should be small enough to fit into
mobile robots and consume very little power so as to
require no cooling and conserve the limited power of
mobile terminals.
The memory units should have very rugged
construction for operation in hostile environments
and even in outer space. Maintenance could be by
unskilled personnel and eventually by remote
service robots. Only solid-state technology is used,
without any moving parts. The outer surface may be
sculptured for better heat transfer and easier
handling by human hands or robots during memory
repair. The input output bus connector may be
similar to the Edison socket now used for light
bulbs. Memory modules may be screwed-in or
removed, like light bulbs, without any additional
mounting hardware. There are two types of memory
units, a "female" unit (CAM), and a "male" unit
(RAM), with the same bus sockets, either male or
female, on both sides. The memory units could be
stacked into larger and larger groups to provide
virtually unlimited storage capacity.
The CAROM memory, shown in Fig. 4, may operate
as a Content Addressable Memory (CAM) or as a
Random Addressable Memory (RAM). Both the
memory address and the stored data are
programmable from the outside by tiny fuses. The
fuses are made either conductive or nonconductive
according to the Autosophy learning algorithms,
where the data determines or creates its own storage
nodes. There is no need for programming or outside
supervision of the internal operations.
Memory addresses are decoded in long chains of
Field Effect Transistors (FETs), where each
transistor is turned conductive or nonconductive as
selected by a pair of fuses. Only one fuse in the pair
is made conductive. A complementary input voltage
is applied through the conductive fuse to the FETs
gate for either conductive or nonconductive. If all
FETs in a chain are conductive, then current will
flow through the FETs to a set of diodes to generate
the output data via programmable output fuses. Only
one chain of FETs can conduct at any one time
according to the learning algorithms. An input
address will therefore produce an output data pattern
as determined by the learning algorithms.
Since only one chain of FETs can conduct at any
one time, the current through the memory device is
constant and independent of the memory capacity.
The voltage input from the driver charges the tiny
capacitance on the FETs gate. The input acts like a
pure capacitive load where the whole memory
device acts like a large capacitor.
The energy used to charge the capacitor can be
recycled through an inductor to achieve very low
energy consumption. When the stored capacitive
voltage is discharged through the driver, in a charge
reversal, then the energy is temporarily stored in an
inductor. The energy stored in the inductor then
recharges the capacitive load in the opposite
polarity. This is similar to an electrical pendulum,
where the energy is being recycled in each charge
reversal. The result is very low net energy loss only
in the conductive layers and in the FETs.
Very small net energy consumption will prevent
heating and conserve the limited power in mobile
robots. Like the human brain, power consumption
will increase with memory activity, such as learning
and thinking. However, power consumption is
virtually independent of the memory capacity. A
very large memory device may have virtually the
same power consumption as a small memory device.
The internally stored energy in the capacitor may
even allow a robot to continue operations after short
power failures.
The memory circuits are printed onto thin foils,
which are rolled into a spool, about the size of a roll
of toilet paper. The spools should only contain the
printed foil as shown in the FOIL LAYOUT without
active components. The driver and receiver circuits
should be housed in a separate bus connector unit in
the center of the spool. Rolling the foil into a tight
spool would provide very rugged and compact
construction. The spool is then inserted into a
capsule filled with self-sealing liquid.
The memory fuses are set by the learning
algorithm whenever a new node is created and
stored in a "next empty" memory location. Each
node may be stored anywhere in the memory device.
Once a fuse is set, from the driver or receiver, then it
will usually remain set to form a read-only memory.
All address bits and the output data bit fuses in a
network node may be set in a single cycle. Fuse
programming will use larger voltages and currents
than normal memory searching operations. Each
column of fuses in a node may be specified in a
"next empty" word selection layer on the foil.
There are several available fuse technologies:
conventional non-volatile memories (PROM,
EPROM, EEPROM, FLUSH) may change over long
time periods and are not recommended. Antifuses
are used in Field Programmable Gate Arrays
(FPGAs), which are more stable and less sensitive to
heat or radiation. Fuses made from electro-active
PLZT (Lead, Lanthanum, Zinc, Titanium) ceramics
may be best for long-term memories. Some fuse
technologies may allow for memory erasure and
recycling of the spools.
Set Fuse 1
There are three basic technologies for
manufacturing the memory spools: Silicon chips
mounted on a laminated foil would provide the
cheapest solution at this time. The circuit shown in
Fig. 4 would be printed on crystalline silicon chips
for separate mounting on the foil. This may yield a
price of about $1 per billion transistors. (The Intel
Pentium 4 contains 330 billion transistors and sells
for $350). The silicon chips need not be error-free
because of the DECAM's self-repair capabilities.
Thin film amorphous silicon transistors and
interconnecting wiring may be printed by vacuum
deposition onto stainless steel foil. This roll-to-roll
industrial process may result in very inexpensive
memory units that can withstand great heat,
radiation, and vibrations.
Printable electronics circuits are printed on plastic
foils by inkjet printing processes. Special semi-
conductive ink can produce the wiring and organic
transistors, diodes, and PLZT fuses.
Both the autosopher and the brain store multimedia
"information" in a saturating omni dimensional
hyperspace format, in which any node may be
located anywhere. Memory repair is far beyond
human intelligence. Repairing an individual
hyperspace memory node is just as impossible as
repairing individual neurons in the brain. Both
systems act like a sealed "black box" to organize and
repair their own memory operations.
Stainless Steel Foil
Set Fuse 2
Set Output Fuse
Figure 4: CAROM Memory using Thin Film FET Semiconductor and Fuses (Patent 5,576,985)
SYSTEMS: Replacing the Data Processing Computer with Self-Learning Machines based on the Autosophy Information
Examples of dual redundant information storage
are found in double ledger accounting and the DNA
helix. In double ledger accounting every transaction
is recorded twice, as a gain and as a loss. Errors in
one ledger can be corrected from the other ledger to
obtain error proof accounting. In biological DNA,
information is stored in two strands wound together
into a helix, where each strand contains the same
information but in a complementary form.
Autosopher store information in two spools, a
male (RAM) and a female (CAM), each containing
the same information in a complementary format.
An error in one spool is automatically repaired from
the complementary spool. Memory repair clears
defective nodes, which are repaired in sleep mode.
Each search or retrieval memory access involves
both spools. In a search access, a MATRIX is
applied to the CAM, resulting in an ADDRESS
where a match was found. The ADDRESS is applied
to the RAM, resulting in a MATRIX, which must be
identical to the original MATRIX input. In retrieval,
an ADDRESS is applied to the RAM, resulting in a
MATRIX output, which is applied to the CAM to
result in an ADDRESS, which must be identical to
the original ADDRESS input. An error in either
spool, confirmed by a "checksum" (A Hamming
code of both the MATRIX and the ADDRESS),
would clear the node for later repair. Defective
memory nodes cleared during normal operations are
repaired in "sleep" mode.
Self-healing involves applying a binary counter
to the CAM to retrieve an ADDRESS and checksum
from the CAM. If the checksum code is incorrect or
empty, then the counter is incremented to check the
next node. Otherwise the output ADDRESS is
applied to the RAM to obtain a MATRIX output,
which must be identical to the counter ADDRESS.
Else the node in the CAM is used to generate a new
node in the RAM in a "next empty" memory
location. Once the CAM has been scanned then the
binary counter ADDRESS is applied to the RAM
ADDRESS. This will result in a MATRIX output,
which is applied to the CAM. The CAM ADDRESS
output must be identical to the counter ADDRESS.
The automatic self-healing facilities can also be
used for rejuvenation and cloning of robot
memories. Removing one spool and replacing it with
an empty spool will cause a robot to automatically
restore the information from the remaining spool
into the empty spool. The removed spool may then
be inserted into a second robot, together with an
empty spool, to produce a robot clone with the same
knowledge and “personality.” Rejuvenation involves
double cloning by removing an old spool and
replacing it with an empty spool. Old robots can thus
be rejuvenated without loss of information.
Considering the enormous head start of the
programmed data processing computer, replacing
the computer with self-learning autosopher will
neither be quick nor easy. However, there may not
really be a choice. Computer technology is
approaching the limits of performance where further
progress will require more and more effort. Higher
and higher speed computation, more advanced
operating systems, and better programming are not
the solution. Shannon's "communication," using
binary digits, is not true communication, and the
programmed data processing computer will not
achieve true Artificial Intelligence. It may be hard to
accept that our entire communications infrastructure
and the programmed data processing computer are
based on a false (Shannon) information theory.
The Autosophy information theory, in contrast, is
based on self-learning networks that grow like data
crystals or data trees in electronic memories. The
intelligence of archiving systems or robots is no
longer limited by the intelligence of the human
programmers. Autosophy may provide an alternate
computing and communication paradigm for the
future, from human-designed and programmed
machines, towards self-organizing and self-repairing
brain-like machines.
Holtz, K., 2004. Autosophy Failure-Proof Multimedia
Archiving. IS&T's 2004 Archiving Conference., San Antonio Texas, April, (2004)
Holtz, K., 2003. Replacing Data Processing Computer
with Self-Learning Autosopher: Impact on
Communication and Computing. SCI 2003 The 7th
World Multi-conference on Systemics, Cybernetics
and Informatics. Orlando FL. July (2003)
Holtz, K., 2002. The Emerging Autosophy Internet.
SSGRR 2002s, L'Aquila Italy Aug. 2 (2002), Paper 140.
Holtz, K. 2001. An Autosophy Image Content-Based
Television System. IS&T's 2001 PICS Montreal
Welch, T., 1984. A Technique for High Performance Data
Compression. IEEE Computer, June (1984)
Ziv, J., Lempel A., 1978. Compression of Individual
Sequences via Variable-Rate Coding. IEEE
Information Theory, IT-24 (1978)
Holtz, K., Langheld, E., 1978. Der selbstlernende und
programmier-freie Assoziationscomputer.
ELEKTRONIK magazin, Germany, Dec. (1978).
Shannon, C., 1948. A mathematical Theory of
Communication. Bell Telephone B-1598, July (1948)