In order to explain consciousness, I want to explain in lay-man's words
the main principles of some of the most important neural networks, being ``error
backpropagation'' (backprop), which, because of publication problems, was
reinvented 4 times, and ``Adaptive Resonance Theory`` (ART), which is more than
30 years old, but only now getting to be valued for what it is worth. So that's old
news, but it's getting time for this news to reach the newspapers. The step from
nervous tissue to the mind and consciousness, and the other way around, is full of
pitfalls, but neural networks are quite boring without looking in that direction.
A neural network is a network of nerve-cells, also named ``neurons'', like we have
in our brains. Neurons have many incoming signals, and (usually) one outgoing
signal which goes to many other neurons. The cells in our cerebrum each have on
average 10,000 entries. Artificial models usually only have a hand full. Cells may
be connected to each other, to sensors (eyes, ears, skin, etcetera), or to muscles
and glands.
Trough its inputs a neuron may be stimulated. When a neuron is stimulated above a
certain threshold, it will become activated, by which it will itself start to stimulate
other neurons. Important is, mathematically spoken, that this is process is
non-linear. The stimulation has to take a certain threshold. In general a large part of
its inputs will have to be stimulated in order for a neuron to become activated.
Special about a neuron is further more that the sensitivity of each of its inputs can
be adjusted by the neuron itself, by which a neuron can teach itself something. The
general principle here is, that if a neuron should have been activated and is indeed
activated, then all inputs that have attributed positively to this will become a little
more sensitive, while the others will become a little less sensitive, etc. OK, you
might say, but how does a neuron ``know'' whether it should have been activated?
And if it knows this, then why should it learn this? The simple answer is that a
neuron learns this for those moments at which it isn't told whether it should have
been activated. In other words, we have to make a distinction between learning and
showing what has been learned. While learning we put a so-called
``teaching-input'' at the output of a neuron. From this teaching-input, and its
ordinary inputs, the neuron knows how to adapt itself.
You can imagine that, when a single neuron has the tendency to produce a certain
output, given a certain input, then a network of neurons as a whole will, when
given a certain pattern as input, tend to produce a certain pattern on its output. In
other words, such a neural network will associate patterns with each other.
Pattern-association, that is automatically coupling patterns which often occur
together at the same time, is indeed one of the things for which a neural network
can be constructed.
My favorite example of a possible use of this is that of an oil-company searching
for oil. This is done by detonating explosives in the ground, and then listening to
the reflected sounds. These sounds are analyzed by experts, and from this it is
decided whether the chance of there being oil is large enough to start drilling. This
is a complex analysis, for which intuition, or at least a clever guess, cannot be
missed. But a neural network is also be particularly well suited for this analysis.
For this we create a neural network consisting of layers. At the bottom we situate an
input-layer, consisting of neurons which are connected tot signals registered by
microphones. Next we have some so-called ``hidden'' layers, and finally, at the
top, an output-layer. Activation-signals run from input layer, through the hidden
layers, to the output layer.
In our example the output-layer only exists of two neurons. Activation of the first
means: ``Yes, there is oil in the ground'', and activation of the second means: ``No,
there is no oil in the ground''. We train the neural network using old, recorded
sounds from the archive, whereby we know through experts or drilling-results
whether there was oil in the ground or not. As such we will make the neural network
associate certain patterns with ``Yes, there is oil in the ground'', and others with the
opposite. Neural networks can in principle learn this such that they also become
very good at deciding from totally knew sound-patterns whether there is oil in the
ground or not--I'll explain this further on.
The adaptation or learning of such a neural network as a whole is of course more
complex than that of a single neuron, but the principles remain the same. A the
output-side nothing changes; there we simply tell the neuron whether it should
have been activated or not. But if this neuron knows this, then it can also tell all
neurons in the hidden layer below whether they should have been activated or not,
etcetera. This mechanism is called ``error-backpropagation'', or ``backprop'' for
short. It has been reinvented several times. Paul Werbos probably was the first, but
they idea became famous through the PDP-group.
Remark that for almost any association the whole network participates because all
neurons influence each other. Because each neuron contributes only a very little
bit to any association, a neural network is able to remember many things even
when some of its neurons die. In other words, knowledge is distributed over all
neurons. More precisely it is distributed over the specific sensitivities between
neurons.
A problem with backprop is that biological systems do not exist in this form. Another
disadvantage is that these kind of networks learn quite slowly. You have to show
them the patterns to be associated many hundreds of times, preferably in random
order. This makes that the network can adapt as ideal as possible to all these
patterns with regard to each other. An accompanying disadvantage then is that
these networks may unlearn old associations through learning new ones.
Besides pattern-association neural networks can also be used to abstract features
from patterns. For this we make a neural network such that we do not have to tell it,
through a teaching input, what the features to be distinguished are, to the contrary,
it will distinguish this itself.
We do this with the aid of neurons which try to switch each other off. In stead of
trying to activate each other, these neurons try to inhibit each other. An inhibited
neuron cannot inhibit any other neuron any more. In the most extreme case only
one winner will remain. Such a winner may then--automatically--provide a
teaching input for underlying neural layers. Remark that, with regard to neural
activation going from input-layer to output-layer, this network is not different from
the layered network described above. But perpendicular to this--that is, within
each or certain neural layers--neurons try to inhibit each other.
A consequence of this neural wiring is that, if the number of possible winners is
limited and smaller than the number of patterns shown to the neural network, then
the network will, in the end, group patterns on distinguishing features. This means
that the winning neurons each will come to stand for certain features of the patterns
shown to the network.
How does this work? Take, to keep it simple, a network which only has inhibiting
neurons in the top neural level. Suppose we randomly show this network two
patterns, and . Let us start with pattern . There will then be a winner
neuron in the top neural level. Which neuron this is, is, the first time at least, pure
coincidence but we know for sure that only one can win, and only one will win.
Call this neuron . Using the principle of backprop again, this winner neuron
next provides a teaching-input to underlying neural layers. As such the neural
network learns to associate with output-neuron . If pattern is different
enough from pattern , then this will, because of this difference, more likely not
activate neuron , but it will activate some other winner neuron, which we may
call . Remark that we create a teaching-input without a teacher. It must be said
that in the simple case it could rather easily happen that both pattern and
pattern activate the same winning output-neuron, but there exist numerous
tricks to prevent this.
Now take a situation where there are more patterns than there are inhibiting
neurons. It is clear then, that it is impossible for every pattern to have it's own
output-neuron. But output-neurons could be coupled to resembling elements of
patterns, such as certain forms or colors, in other words (general) features or
properties of these patterns. A structure like this will come into existence
automatically because, not only can there be only one winning neuron, there also
always will be one winning neuron, and if we keep showing all patterns to the
network randomly, then it will keep adapting until everything fits best.
If we allow for more than one neuron to win, then certain combinations of
output-neurons will become activated, which together form an analysis or
deconstruction of the input pattern. Wherever we, human beings, look at, it will
immediately decompose in a number of parts. I refer to this as an attention pattern.
The principle of inhibition can be build into any neural network. It automatically
makes networks a little more ``intelligent''. It makes that a neural network structures
itself, by finding the most important features, organizing itself arround these, and by
forgetting the less important details more easily. This is abstraction in a very
abstract sense--I call it immanication, with immanence being the opposite of
transcendence. It will, for instance, also help with regard to the example of the
oil-company mentioned earlier. In our own brains there are many neurons which
seem to have functions dealing with this.
A problem with backprop is that inside our brains error-backpropagation cannot
work that way. Over 30 years ago Stephen Grossberg invented a
mechanical-mathematical principle which does not have this problem. He calls his
theory ``Adaptive Resonance Theory'', or ``ART'' for short, and he explicitly
connects the principle of resonance in his theory with consciousness. In completely
different words than he uses I would like to try to explain his theory. In my
explanation I, further more, will refer to the human brain and mind. ART as
Grossberg describes it is logically and mathematically sound, but it is too detailed
and complex for me to show you some of the more general issues I want to address.
The fact that Grossberg brings mathematics and consciousness together in one
scientific theory remains magnificent, whatever one may think of the contents of it.
The fact that he wrote down his theory, and built his neural machine, so long ago,
and only receives recognition know is, I think, sad, and in my opinion it is shows
how science works in daily practice when we are dealing with truly knew
developments.
Take two neural networks, each consisting of about seven neural layers, and lay
them on top of each other, but in opposite direction with regard to each other's
input and output side. So, layer 7 lays on top of layer 1, layer 6 on layer 2,
etcetera. Neurons are, in between each of these networks, connected one to one
such that they tend to sensitize each other when activated themselves but without
them being able to activate each other. A such two such neurons, each from the
other neural network, in fact together form one single neuron--in the explanation
of Stephen Grossberg it is indeed exactly this. The result should be that the two
networks tend to mirror each other, but with the neural activation flowing in the
opposite direction with regard to each other.
Figure:
A hierarchical neural network split in two halves
The idea is that we can perceive, abstract, and deconstruct things through the first
neural network, while we can imagine and construct things through the second
network. For this we connect the input-layer of the first to our senses. This network
I call the deconstruction neural network. The output of this network I connect to a
winner-take-almost-all neural layer consisting of mutually inhibiting neurons. This
winner-take-almost-all layer actually functions as an attention mechanism. This
neural attention mechanism does no more than to allow for a small number of
winners to stimulate the second neural network.
So the winning neurons of the attention mechanism deliver their output on the input
of the second network which I shall name the construction neural network. The
attention mechanism activates a small number of neurons in the construction neural
network. From this the activation flows back in the direction of the senses,
something which we experience as imagination when we do not actually see the
thing imagined.
Attention-like mechanisms exist in every neural layer, but in lower neural layers
there is not a single attention mechanism, instead there are hundreds to hundreds
of thousands, each of which cover a small area in the neural network. Their
function is abstraction. Imagine: You look at a painting. That is a complex pattern. It
enters your neural network through your eyes. Inhibiting neurons bring about
abstraction. In the first layer lines, colors, and simple movement is detected. In
higher neural layers ever more complex forms are detected. In the end the attention
mechanism only allows for five winners. From these five winners neural activation
flows back through the construction neural network, in the direction of the eyes.
Because the deconstruction and the construction neural network tend to mirror
each other, the back-flowing activation will tend to be a (re)construction of what is
perceived. When this succeeds the network enters into a state which Grossberg
calls ``resonance''. Resonance is a relatively stable situation of the neural network
which, given the properties of neurons, will automatically lead to strong learning for
the simple reason that neurons, in this state of resonance, will stimulate, or not
stimulate, each other for a relatively long period of time. In a human embryo the
directions of activation will be rather random, and the accompanying resonance will
therefore be little, but there too the activation will be directed such that the two
networks tend to mirror each other. As such we form ourselves an imagination along
with what we see. In other words, in this way something may attract our attention.
The other way around, because the constructive neural network influences the
constructive network in exactly the same way, our imagination of something will
also make us attend to something. In this case we first, as it where, have an
attention pattern, and from this choices in what we see are made. Along both
routes, but especially along the latter route, a certain point of attention will lead to
another point of attention, again and again. This is further more helped by the fact
that specific attention naturally is quickly exhausted. As such we have one basis
for `thinking' which humans and other mammals have in common.
Remark that the rather transcendental information which enters my eyes--being
just a bunch of colors--is transformed into the activation of only a small number
of neurons, which together form a pattern of attention. So attention in fact is the
summit of abstraction. The mirroring between the constructive and deconstructive
neural network thereby is a biologically possible way of error-backpropagation,
because mirroring in fact entails that every difference in activation between the two
neural networks is an `error'. This error is detected and corrected in both networks
at neural level, and all this solely on the base of mechanical principles.
Although my explanation is completely different from that of Stephen Grossberg, it
is, I think, largely a summary of his theory, with the addition of ideas of my own.
Grossberg has proven that ART neural networks work very well, learn much faster
than ordinary backprop networks, and do not suffer from massive loss of memory.
With regard to one thing I disagree with Stephen Grossberg. Grossberg couples
consciousness to resonance. I think consciousness is impossible without
temporality.
With regard to our imagination many things play a role. Especially time is very
important. We imagine ourselves things by imagining them in time. Our attention
shifts from one aspect to another, and we do not do this without structure. Therefore
this temporal structure must, at least in some respect, be laid down in a
non-temporal structure. Attention plays an important role in this, because in the
attention pattern successive matters can be associated with each other. More
specialized planning- and language-mechanisms are equally important, especially
for humans. All in all the middle and higher neural layers make for a coherent
structure in our thoughts and imagination. They form our ``pre-consciousness''. Our
actual consciousness only exists in time. That we have the illusion to be conscious
of much more is simply caused by the fact that we are not conscious of that of
which we are not conscious. We only are conscious of the color of the eyes of a
squirrel at the moment that we are conscious of the color of the eyes of a squirrel.
A nice example, and also a metaphor, of temporal images, and ones the reason for
me to discover all this, is the working of a movie or television program. Watching a
movie, every 2 to 5 seconds a new attention pattern arises--you may have the
illusion to see more, but close your eyes and count what you saw. At numerous
moments in time such patterns are connected in the here-and-now, for example
when a certain person, or a certain house, again attracts our attention. A movie is
an image in time, that is a temporal image, but these kinds of association based on
recurrent elements also form a more permanent image from which we can remember
things. If associations and meanings are woven into each other well enough, then
lower neural layers will develop association patterns too. Our consciousness is
much like such a movie. Our consciousness is an image (a `knowing') in time.
An important form of mirroring which I, be it in completely different words, have
addressed is that which in semiotics, semiology, and linguistics is named
``reference''. Here I refer to the mirroring of the world inside our neural network. For
instance, the word ``cow'' refers to a cow. The activation of each neuron in an
attention pattern (usually) refers to something in the world. Our neurons are small
mechanisms which are part of the world, so neurons work and are being worked on
solely according to `physical' principles. Memory in the form of adaption of
sensitivities between neurons play a fundamental role in this mirroring. The mirroring
of the world from moment to moment into a neural network--i.e. reference--is in
itself of little value. Important is that a structure in time with regard to these
mirrorings is remembered, such that it may later lead to comparable structures in
time within this mirroring matter, being the neural network. In saying this we go
much further than `physical principles'. Not that it is in contradiction with it. It just
becomes another topic because memory and structures in time come to play a part
in understanding it. Not only do we have to think and speak at another
aggregation-level because we would otherwise not be able to see the threes
through the wood, the system at this level also is a reality in the sense that this
cloud of molecules which we call a neural network or brain is a structurally
relatively stable situation which as a whole interacts with its environment. The fact
that it acts as a whole is due to the structure as described above. As such it is
indeed the case that `the content' of the neural network, such as you or I myself, is
an actor in causal relations.
If we make a couple of other, big leaps, connecting muscles and bones to our
neural network, which are driven by our imagining of our movements (in other
words, they are coupled to the output of a special part of the constructive neural
network), and further more introducing some clever planning- and
energy-mechanisms, plus instincts and an underlaying reptile brain, then we have
something which qua intellect is already much like an animal. To create a human
being we have to explain a little better how language and self-consciousness
works.
Everything exists in time. Nothing special about that. But in what sense plays
something which happened earlier a role in a later situation? You might say: I have
a memory, and because of that I remember things, but what does that have to do
with consciousness? Then I say: OK, but you keep speaking about ``I''. Formulate it
without ``I'' and make it into a machine, because that is what I need to explain
consciousness. Without ``I'' one speaks of a more timeless memory, and a more
temporal imagination, i.e. a more temporal consciousness. So, an important
presupposition for me is that consciousness cannot stand still. I cannot imagine
anything with regard to a consciousness that is frozen, but neither can I imagine
consciousness without a memory. Of course memory changes too, through
learning, but it does so more slowly.
Suppose a leave whirls on the street from point A to point B. Next it whirls to point
C. The fact that it first lay on B is important, because otherwise it would probably
never have landed exactly on C. This is a kind of memory, but as soon as the
leave lays on C there is no memory of A and B anymore. One peculiarity of
consciousness is that is does leave a traceable trail of its own ongoing, and it even
does this such that the elements of this trail will return automatically ever again
when necessary. The nucleus of my story is in the relation between this
remembered trail trail and the ever changing content of consciousness. This holding
on I also name ``mirroring''. ``Mirroring'' in this sense does not refer to the mirroring
taking place between the constructive and deconstructive neural network. It refers
to the mirroring of the world into the neural network. The result of this mirroring is
what one tries do describe in Artificial Intelligence theories (AI), but they do little
with the mirroring itself. In contrast for me it is essential to understand this mirroring,
that is the relations between the timeless mirroring, the world, and
consciousness--the latter two both being temporal.
Why is this an explanation for consciousness? I think you first have to grasp that
my idea actually is a much too simple explanation for any sensible individual.
Secondly my idea is terribly abstract. There are potentially many possible
implementations, even though it is quite difficult to come up with examples. Thirdly,
consciousness as mentioned above in fact is only interesting if it is a
consciousness of something. For this one needs something like a neural
network, but my definition of consciousness is in fact much more abstract.
An example of a consciousness which is not a consciousness of something
is given in Hofstadter's book ``Gödel, Escher, Bach'' in the form of Aunt Hilly, that
is an ant-hill. Ants, in their patterns of movement, form certain repetitive patterns,
because they follow each others trails and signals. These patterns are influenced
by their environment, and an ant-hill does have a tendency to preserve itself, so
so consciousness of something cannot be denied to it, but what does this entail qua
content? Almost nothing. As long as we realize what the content of this
consciousness entails it is not strange to attribute consciousness to this.
This said it may be a question whether one should call a consciousness which is
not a consciousness of something ``consciousness''? Maybe not, but I like to
resolve this by pointing out that systems which do not have consciousness of
something usually are so very simple qua consciouness that the question is not
important at all. The only exception might be AI, but here too it must be said that the
topic of AI is just that which happens in the very top level of my system.
Did I answer the question of whether this is an explanation of consciousness? I did
if you agree with my definition and description of consciousness. In an abstract
sense my description of consciousness is in agreement with my idea of it. Since I
have also been able to describe the mechanism of it, I have explained it. The idea
is so very abstract that it does not seem simple anymore, but it is.
To this simple idea of consciousness one can next put thousands of questions,
which all need to be checked. Take the importance of the sensibility of thoughts.
Certain mental illnesses make people so confused that their consciousness is
clearly limited by it. In a psychoses thoughts can be like a whirling leave, and this
can destruct memory and consciousness considerably (desintegration). So
sensibility has everything to do with coherence and structure, and this has
everything to do with consciousness because if there is coherence in the timeless
``memory'', then the consciousness arising from this may have this coherence too.
But if this coherence is too strong, then it may become suffocating. That is exactly
the property of neurosis (fixation).
Another important aspect of my consciousness-thesis is that I want to deconstruct
the illusion that we are conscious of many things at the same time, which is what
most people seem to think. Consciousness at a certain moment is totally senseless
to me. This is more an assumption than something a can prove, but I did want to
explain how it seems to us that we are conscious of much more on each moment
than we in fact are.
No consciousness is fully without self-consciousness because of its relatively
timeless ``memory''. The past of one's consciousness always plays a role in one's
future consciousness. Only the way in which, and the directness with which one's
consciousness participates in this can differ greatly. Human beings can
systematically register their own consciousness (the mind's eye) using mechanisms
which form the foundation of language, and language itself. Animals associate
events with their own consciousness (through their attention patterns), but they
cannot really do this on purpose. They cannot actively associate their contents of
consciousness other than through their actions because an earlier content of
consciousness is for them largely gone the moment the next is there, with the
exception of the timeless memory of this, of course. That would have been the same
for humans, where it not that we can use language to manipulate our own and other
people's consciousness. This said, for both humans and animals successive
attention patterns will usually overlap, and therefore there will arise associations,
either directly or indirectly, in the same way we make them when watching a silent
movie. I think animals differ considerably in their talents. Mammals and birds can
certainly perceive and remember temporal patterns (in the basal ganglia and/or
prefrontal, motor cortex) and through this develop a certain self-consciousness like
we, humans, have. I cannot distinguish anything like this in Aunt Hipe.
In my final definition of consciousness I left the importance of attention out. The idea
is that one does not need this if one asks not whether or not something has
consciousness, but what this consciousness looks like. This said, I cannot imagine
how there could be a coherent consciousness in which everything comes together,
and in which consciousness could gain a respectable structure, without attention.
As an admirer of Carl Popper I should therefore add attention as essential to
consciousness, since he would say that I should make this into a hypothesis in
order to give others the possibility to falsify it. this said, I do know a replacement for
the attention mechanism, namely instinct--think of cats--but in fact instinct is a
kind of attention too.
Attention makes that even the strongest atheist in the end tends to arrive at a
single, most important, ordering principle. In as far as someone does not succeed in
this, his thoughts, or even his personalities, remain like lose sand, for instance with
ethics detached from scientific or practical thinking, which is quite wrong in my
opinion.
Particularly intriguing with regard to consciousness are qualia such as pain, and
feelings in general. Understanding these is complex partly because old
brain-structures and inborn reflexes play an important role. Instincts are probably
connected to our newer brains differently than our senses and muscles. The
connections are non-specific, by which I mean that an instinct stimulates many
neurons. Through this instincts can globaly change our perception. Suppose, for
instance, that a newborn baby has the instinct to activate some area of the neural
network whenever it sees two darker spots. These darker spots will usually be the
eyes of mother or father. As such the human neural network will tend to save
everything that father and mother do at a certain location. The other way around the
instinct can later ``use'' this information, through activating this piece of neural
network. We could analyze this as being our ``superego''.
As such instinct is a kind of inborn, vague attention. A particularity of pain is that it
is at the same time a very strong attractor of attention, and also enormously
stimulates the neural network as a whole. Although pain itself is not really
remembered as such, longterm pain must lead to restructuring of the content of the
brains. It is anyway very fatiguing. The temporality of pain is in the fact that it does
not stop and that it cannot be ignored. At the same time, through its attraction of
attention, is also leads to a narrowing of consciousness. Fear seems to work more
or less the same except that there is no other source of pain other than a thought.
A part of the experience of qualia seems to lay in the interaction with older
structures of the brain. It is anyway something which we share with many animals.
The only, but very distinguishing, particularity of humans is that we can imagine
much more things too ourselves, and thereby we can experience much more fears.
As such an animal farm is not always a concentration-camp, but vivisection
remains silly and scoundrelly--even if for example cats also like to play with the
death of their victims, and even if the weighing of interests is often difficult. A
human being who panics is qua consciousness not much different from an animal in
panic, and hunted and brutally overruled animals do not behave very different from
hunted and brutally overruled humans. This is exactly what makes this profession of
brain- and neural network studies is so rotten to me; there are so many brutes
involved in it.
This said, we do, of course, not live in Paradise. We have to survive too. Just
imagine having a manic-depressive daughter and hoping for some medical wonder
to cure her--not that I expect scientists to really have other motives than
curiosity, fame, and food, but it still remains an argument.
I want to try to give this story a nice ending... I miss something in my exposition of
how feelings work. Maybe this is only the missing itself; the fact of knowing that I do
not know everything; the hole in my mind, and the desire for something which I do
not have, if even a solution to get rid of my pain. But something tells me there is
something else. Whether I think of toothache or feelings of beauty, there seems to
be more. The curious thing is that I experience this with regard to earthly matters.
Consciousness and individuality; that I do understand, but... that feeling of
community, and the both peaceful and crude existence; that I do not understand. It
seems as if this, which I do not understand, is all that which enters my
consciousness, but which lays outside my power, and outside my imagination. It
steers and pushes me, and I can go along with it or resist it, or think of a trick to get
it within my power anyway. It is fantastic and magnificent.
To contents of this article are addressed more deeply in an Internetboek by the
author entitled: ``The Time Machine, Prototype of a Conscious Machine'', which
can be found at http://tm.bedenkerij.nl/
Publications of Stephen Grossberg can mostly be found on the Internet too,
unfortunately quite often referring to vivisection, but for the nucleus of his theory
this is absolutely superfluous. See http://www.cns.bu.edu/Profiles/Grossberg/
James Anderson and Edward Rosenfeld wrote a beautiful book in 1998 about many
the people involved in the development of neural networks: ``Talking Nets. An Oral
History of Neural Networks'', The MIT Press: Cambridge, London.
and 
. Let us start with pattern 
. Using the principle of backprop again, this winner neuron 
. Remark that we create a teaching-input without a teacher. It must be said
that in the simple case it could rather easily happen that both pattern 