Abstract

	
Gradual uploading could be considered as an alternative to neuron-by-neuron
uploading. In this article we discuss some technological aspects of this 
problem, including computer-brain, brain-computer interfaces and necessary 
computational power for mind simulation. We interpret the uploading system
as a fault tolerant neural network (where damage would be brain death), as 
a "black box" (brain) imitation system. We use children mind evolution model 
for the description of uploading system evolution, where the external world 
is changed to the brain world, receptors - to the brain-computer interface, 
motor functions - to the computer-brain interface.
Speech communication as an interface prototype is discussed.

Contents

1. Introduction

2. Functional uploading

3. Interpretation of gradual uploading

4. Brain - computer interface.

5. Computer - brain interface.

6. Computer

7. The system of brain cognition

 	 7.1 The principle of world modeling

	  7.2 Comparison between real world and brain cognition.

	  7.3 Evolution of the external system

8. Speech as an interface

9. The problem of person identity

10. Discussion

11. Conclusions

12. Acknowledgments

13. References

1. Introduction

The gradual uploading of mind (also known as gradual extension, metamorphosis
[7], soft uploading, brain enhancement) could be evaluated as an alternative
of atom-by-atom, or neuron-by-neuron uploading.

	The gradual uploading requires addition of artificial neural network (or other 
means of AI) to the brain. After this, the brain and the external system would 
operate as one system. During this common coexistence, the memory and other 
functions of mind would gradually grow into the external system and would 
survive after the brain's death. Therefore, the death of the brain would not 
be fatal for the person, i.e. main goal of uploading would be achieved.

	The goal of this article is to discuss some ideas, how this problem can
be solved, what is going on in this field of research, and what is necessary 
to do in the future.

The gradual uploading would be uploading of mind functions, not uploading of a 
morphological brain architectures.

	At first, the morphological structure of brain is defined by not only 
functional necessities. Evolution is very stochastic process and accidental 
factors contributed a lot to the brain structure.

	At second, some functions (homeostasis, most motor reflexes etc.) are 
not useful for the cyber system, if the 'cyber-body' differs a lot from the 
human body. In addition, it looks like those functions are not so important 
for the self-perception.

	The mechanisms of brain functioning and functions unimportant for the 
self-perception can be abolished, adapted or changed. This standpoint could 
greatly facilitate the solution of the uploading problem.

External neural network of the uploading system could be implemented using
several ways. The interpretation of performance of this network can be various
too:

1. The neural network functions are distributed over the network (over large 
field of network, at least). In some cases, the neural network is a fault-
tolerant i.e. after removal of some neurons, survived part of network can 
inherit almost complete collection of parent network functions. Under certain 
conditions similar properties could be used for the separation of external 
system in case of death of biological brain.

2. The brain can be interpreted as a "black box" with complex inside structure
and functional properties. If the sufficient quantity of information about 
brain properties were obtained, the imitation or reconstruction of these
properties would be possible.

3. Some ideas about organization and interpretation of external neural network
can be supported using the existing example of powerful neural system - the
human brain. Evolution of external system could be interpreted as a child
evolution, in this case. Some cognitive science, psychological and pedagogical
models can be used for this evolution description. We will discuss this point 
of view in details later.

	In any case, the problem of technical implementation of gradual uploading can 
be divided into:

	external system software problem,

	external system hardware problem,

	interface computer - brain problem,

	interface brain - computer problem.

There are several possibilities of creation of this interface. One way is to 
implement some electrodes or microchips into the brain. At this time some 
examples of this interface are available [9,10,11 ]. Main disadvantage of this 
method is necessity of surgical invasion into the brain.

	Another way is to use multichanell electroencephalogram (EEG). Some 
implementations of EEG brain-computer interfaces are also available at this 
time [1]. For the EEG, the minimal distance between electrodes must exceed 1 
centimeter, and this factor limits a number of electrodes, i.e. EEG interfaces 
permeability. The EEG could represent brain information from the cortex only. 
In addition, the EEG channel can represent only a sum of activities of many 
neurons under channel electrode.

	Other possible technology for the interface could be magnetoencephalography,
using SQUID (Superconducting QUantum Interference Device) [13]. This is a most
sensitive technology for registration of magnetic fields, at this time.

	For the gradual uploading goals, the brain-computer interface should have some 
specific features. One problem is representation of necessary brain information 
in the cortex field under EEG, MEG electrodes or near implant. This problem 
can be partly solved optimizing the location of electrodes according to our 
current knowledge about cortex structure.

	More powerful method would be a brain adaptation itself. The brain - computer
interface would have feedback coupling (computer - brain interface), and this 
fact would ensure good co-operation between brain and uploading system.
Through this co-operation, the brain adaptation would happen itself. Besides,
some additional means could be implemented for better brain adaptation and
those means could create necessary ways to transfer the information into the 
brain. As one of examples of this adaptation some human learning methods can 
be considered which relates the artificial electroactivity in the deeper fields
of brain with the electroactivity in the fixed field of cortex.

Some examples of human brain adaptation (Braille reading method, the speech of
gestures etc.) show that such adaptation can be possible in principle, though 
these possibilities diminish during the human life.

The brain implants could be used for information translation to the brain 
too [9,10,11]. Other way is to use sensations of human body for the 
transmission of computer - brain information.

Human body uses many nerves to translate information about external world to 
the brain. Some of nerves could be used to translate the information from the 
computer to the brain. One way is to use tactile nerves. In this case, the 
interface hardware would include a matrix element, each consisting of a small 
vibrator (like small needle).

The tactile brain input was widely investigated for hearing disorder 
compensation using tactile sensation. Some bibliography about this studies can 
be found in [2,3].

	Another way to transmit information transmission is to use visual sensations.
Such interface would have better information permeability, but it might
interfere to world`s visual perception.

Using connection machine  the processing speed up to  1.3 * 10^9 synapses per 
second can be achieved [6], and using transputer system  the processing speed 
up to  2.7 *10^9 synapses per second can be reached [8]. In principle, the 
most powerful current parallel hardware (1.8 *10^12 FLOP) [14] could implement 
up to 10^10 - 10^11 synapses/second.

	The evaluation of the computational power of human brain very uncertain at 
this time. Some estimates of brain power could be based on the brain synapses 
number and neurons firing rate. The human brain have a 10^11 neurons and each 
neuron has average of 10^2 - 10^4 synapses. The average firing rate of brain 
neurons is about 100-1000 Hz.
As result the brain modeling would require the computational power
of
 10^11 neurons * (10^2-10^4 synapses/neuron) * (100-1000 Hz) =
10^15 - 10^18 synapses/second.

	Other estimations [12] show that the brain power could be between 10^13 and
10^16 synapses/second.

	On the other hand, some factors could diminish the necessary computer power.
Most artificial neural networks (backpropagation, Hopfield, Kohonen, Grossberg
etc.) use  neuron activity represented by analog signal amplitude (firing
frequency modulation), not a firing signal. Computer simulation in this case 
is more simple and speedy.

	Change of all brain neural network to amplitudical network could be problematic.
For example, the temporary pattern of integration/summation is complicated for
amplitudical neural networks, but more usual for firing neurons. But some
systems (spatial processing parts of visual system, for example) could be
changed and simulated more effectively.

	The brain have many systems (sensors and motor channels, homeostasis
system) which are not that important and could be replaced after uploading. Many
sensory - motor tracts would be incompatible with new body properties, many
old reflexes would be useless etc. The large sound articulation control network
could be replaced to few transistors, for example. The similar changes could 
be done to most sensory-motorics systems of the brain, and this would let 
decrease the necessary computer power.

	The brain neural networks are redundant. Besides not all neurons in the
brain are active at the same time. Those facts could lead to decrease of the
necessary computer power.

	The evolution of external system will take some time (the children evolution
take some years), and during this time computer will become more powerful 
(child brain after birth has little number of synapses too, but during his 
evolution this number increases).

Born child has sensors, effectors, motivation, sufficiently powerful neural
network (which has necessary architecture for self-organization). Via sensors 
and effectors child has contact with the environment (physical and social). 
During the children evolution, brain creates internal model of external world. 
Most of intellectual functions of any person can be interpreted  as 
information operations in that model [4]. The same principle can be realized 
in the cyber system too, if the system has all necessary attributes (contact 
with environment, motivation etc.). Some studies in the fields of neural 
networks and AI [15,16] show how model of external world could be formed.

The external world of the cyber system could not only be a physical world.
If the sensors and effectors of cyber system were changed to brain - computer 
and computer - brain interfaces (i. e. the system would be connected with 
human brain "world"), the system would create an internal model of this brain 
(like as model of the physical world).

	The "brain world" would not be very different from the real world.
Some similarities:

1. The real world cognition is implemented via sensations.

The brain cognition would be implemented via external system sensations -
   brain-computer interface.

2. Brain have a possibility to change the real world.

 The external system would use the signal to the brain (feedback coupling)
   for the brain change.

3. For the real world description the brain use the sensoric patterns, symbolic
   categories, combinations of patterns and categories, relationships between 
   patterns and categories.

  For the "brain world" description would used patterns from brain-computer 
   interface, symbolic categories, combination of these, relations between 
   these patterns and categories.

2. Many uncertainties exist  in the information from the real world.
   Many uncertainties exist in the information from the brain.

On the other hand, there are some differences in the case of external system:

1. There are only indirect possibilities to influence the brain evolution
   process. In case of the external system it is possible to influence evolution
   of every single neuron.

2. The real world cognition is formed from two main parts: sensoric cognition
   (unprocessed information) and symbolic cognition (this information are
   processed during society evolution). Very roughly, these could be interpreted
   as neural network learning with supervisor: the sensor information would 
   be neural network input, the symbolic information (notions, for example) -  
   learning target.

 In the case of external system, the "sensor" and "learning target" inputs
   would be received from the same environment - the human brain. The brain 
   information contains not only unorganized sensor information (like as 
   physical world), but the more organized, abstract information too. This 
   information could substitute the symbolic information in the case of real 
   world.

In the cognitive level, the evolution of this external system will depend
from initial and latest motivation of the system. The control of this
motivation would allow to control an evolution of the external system.
Inborn  motivation is a very simple: stress or satisfaction. Genetic defines 
only which sensation is stress and which is satisfaction. The child evolution 
up to about two years could be described as a sensor stage. In this stage 
most sensors categories are memorized only. The motivation of the child are 
reference to these created categories.

	After sensoric stage the semantic connections between categories are created.
Through self-organizing processes and social influence (speech mainly) the
abstraction process of categories starts. The latest motivation of child is 
reference to new, more abstract categories.

	The external system could implement similar model. It is necessary to have some
initial ("genetic") motives for the evolution of this system. Because the 
receptors of system would be connected to the human brain, the initial motive 
could be defined as some certain state of the brain, for example satisfaction. 
This state must be defined before system initialization, i.e. is necessary to 
know what brain output the satisfaction has. This properties could be defined 
subjectively or objectively, for example using some psychological or biometrics 
tests (like false detector).

	In the sensoric stage the external system would create some categories of
"brain world" and use these categories to define further motives of system
In this stage, the external system have to reply clearly to the concrete brain 
state.

	Later, the categories of the "brain world" would be related via semantic
connections. Symbolic brain information would be used for category
abstraction processes. During those processes the motivation of the system
could be improved, using these new categories.

This is a very rough description only. Real processes would be much
complicated. An external system abstraction, rule extraction functions, and,
the most important the human brain adaptation must be investigated.

	The possibilities of operation of external system after brain death could
be supported by hemispherectomy - human brain hemisphere cortex removal in
the tumor case. After the removal the brain`s ability survives, though some
specifics hemisphere functions could be lost [5].

It is necessary to plan, how the external system will contact with external
world after brain`s death. This problem could be solved using artificial
sensors and effectors, artificial means of initial processing of sensor
or motor information, which would connect the external system after the 
brain`s death. Another way is to organize the sensor and motor channels 
during system evolution. In this case it is necessary to plan, how the 
additional external world sensation would influence the extension system - 
brain evolution.

The example of the gradual uploading rudiments could be speech communication
between persons. The human speech could be interpreted as a interface
between two brains, though it does not have the necessary rate for uploading.
At average, human being uses about 10^4 notions - words of the speech.
The speed of speech is about 2 - 5 words per second. The word information
capacity is about about 20 bites, roughly (1 word from 10^4 is 13 -14 bites 
plus 6 bites of grammatical information). So, the speech information 
permeability would be about 40 - 100 bites per second.
In spite of low speed, the communication results are very good. Large part
of our knowledge is acquired by using speech or reading. Sometimes, after long
contact with other person we can observe phenomena, which could be interpreted 
as rudiments of uploading or merging.

The very important goal of gradual uploading is the subjective perception of
continuity of person, during all evolution of system.

	It is possible, that the brain and the external system could work as one system. 
The fighter pilot perceive machine like part of his body. The external system 
would have more communication channels and would have longer contact with the 
brain (it would be best if that contact could be unbroken). Therefore, the 
subjective self-perception of person could be not only be one's brain 
perception, but the common brain - external system perception.

	During the gradual uploading the common brain - external system neural system 
would be dramatically changed and these changes could be interpreted as person 
identity change. But these changes would have the evolutional character during 
the time of gradual uploading procedure.

	During some years the person neural system will be changed in any case (with
or without gradual uploading). Some new impressions, emotionality, motor
skills etc. would be acquired, some - lost. But these changes have the
evolutional character and this fact assures the subjective perception continuity
of person identity.

	The most drastic leap would be death of part of common neural system -
biological brain. Fortunately, the human brain have unconscious states like
sleep. During these states the external system could continue her activities.
When the brain would wake up, these impressions could be interpreted like as
"impressions after brain`s death". Such experience could decrease the
psychological problems.

Such person identity perception mechanism would be unique property of gradual
uploading. In the neuron-by-neuron uploading case this problem can arise
much more strongly.

It is necessary to mention some problems necessary to solve for
gradual uploading implementation.

1. Although some examples of world modeling is created at this time, is
necessary to perform many investigations for the creation of brain power 
cognition system.
Necessary computer power for mind simulation is unavailable too, though
the difference is not very large.

2. For the external system creation it is necessary to create some
conception of what brain functions should be uploaded to the external system,
what functions could be imitated and what functions wouldn't be necessary
for the person identity and functioning.

Meantime the same problem would be faced in the atom-by-atom or neuron-by-neuron
uploading case.

3. Very important problem is the necessity to create information transmission
channels into the brain. Some methods need to be created for this adaptation.

4. Other problem for the interface is brain-computer and computer-brain
interface's permeability. For good brain - external system co-operation we need 
to have sufficient information exchange rate. The question what rate is 
necessary, and what rate the interfaces would have, could predetermine 
usefulness of these interfaces.

Although the gradual uploading problem is a very complicated, many components
of gradual uploading systems could be available at this time.
Some of them must be adapted for these purposes, some of them must be improved,
but there is no need for any technological qualitative leaps for its 
implementation.

	I greatly appreciate very useful discussions with Anders Sandberg and Joseph 
Strout, which gave me a better understanding of this problem.

1. http://www.aleph.se/Trans/Global/Uploading/richard.seabrook.brain.computer.interface.txt

2. http://mambo.ucsc.edu/psl/tactile_speech.txt

3. http://mambo.ucsc.edu/psl/speechreading.txt

4. Marvin L. Minsky, Matter, Mind and Models, "Semantic Information
Processing," (Marvin Minsky, Ed.) MIT Press, 1968

ftp://ftp.ai.mit.edu/pub/minsky/MatterMind&Models

5. Floyd E. Bloom, Arlyne Lazerson, Laura Hofstadter, Brain, Mind and Behavior,
New York, 1985.

6. Alexander Singer, Exploiting the Inherent Parallelism of Artificial Neural
networks to Achieve 1300 Million Interconnects per Second, International Neural
Networks conference, Paris July 9-13, 1990

7. Thomas Donaldson, Metamorphosis An Alternative To Uploading, Cryonics, May
1990

8. H.P. Siemon, A. Ultsch, Kohonen networks on transputers: implementation and
animation, International Neural Networks conference, Paris July 9-13, 1990

9. http://www.ee.surry.ac.uk/Personal/D.Banks/devel.html

10. http://aramis.stanford.edu/cis/research/LabProject94/StanfordDVA.html

11. http://atlas.axiom.net/cornucopia/chips.html

12. http://www.merkle.com/brainLimits.html

13. http://www.hypres.com/~masoud/digital.shtml

14. http://www.ssd.intel.com/press/tflop.html

15.Uwe R. Zimmer, Robust World-Modelling and Navigation in a Real World,
NeuroComputing, Vol. 13, Nos. 2-4, 1996

gopher://ag-vp-gopher.informatic.uni-kl.de/44ftp%3aPublic%3aNeural_Networks
%3aReparts%3aZimmer.Robust.ps.gz

16. http://isd.cme.nist.gov/brochur/SPWN.ps