This paper has not been completely converted to html. The symbols in the mathematical derivations are consistant, but may appear strange. I will rectify this as soon as possible. In the mean time I believe that it is comprehensable and it is better to make it availabe

Artificial Neuroconsciousness: An Update

Igor Aleksander Department of Electrical and Electronic Engineering, Imperial College London, UK


The concept of a theory of artificial neural consciousness based on neural machines was introduced at ICANN94 (Aleksander, 1994) [15]. Here the theory is developed by defining that which would have to be synthesized were consciousness to be found in an engineered artefact. This is given the name "artificial consciousness" to indicate that the theory is objective and while it applies to manufactured devices it also stimulates a discussion of the relevance of such a theory to the consciousness of living organisms. The theory consists of a fundamental postulate and a series of corollaries. In this paper the series of corollaries is extended and illustrated by means of characteristic state structures. Studies of artificial neuroconsciousness aim at two results: first to provide a single perspective on many mechanisms which perform cognitive tasks; and second, it provides an explanation of consciousness which stands alongside the many discussions found in the literature of the day [1-4].

1 Theory

The theoretical framework used in this work has one fundamental postulate from which follow 12 corollaries. This framework has been inspired by Kelly's [5] theory of "personal constructs"which explains the causes of personality differences in human beings.

The Fundamental Postulate: Consciousness and Neural Activity.

The personal sensations that lead to the consciousness of an organism are due to the firing patterns of some neurons, such neurons being part of a larger number which form the state variables of a neural state machine, the firing patterns having been learned through a transfer of activity between sensory input neurons and the state neurons.

The words of this postulate are intended to have specific meanings which need to be stressed so that the corollaries which follow should make sense.

Personal sensation: Much of the controversy surrounding consciousness comes from the problem of infinite regress. Here it is implied that neural activity leads directly to personal sensation so dismissing the problem of infinite regress.

Firing patterns: Neurological terminology has been adopted to refer to the output activity of a group of neural elements. In an artificial system 'firing patterns' could refer to any measurement of the output quantity of the elements which constitute that system.

Neurons: This adoption of this neurological term is used to indicate that the theory is that of a cellular system where "neuron" is the name given to a basic cell.

Neural state machine: A state machine is the most general model of a finite computing process - it calls on the concept of an inner state which is a function of input sequences. Neural versions assume that neurons generate the variable values which, when taken together, form a state. (Corollary 1 formalises this notion and the generality of neural state machines has been argued elsewhere [6] ). Learned: Neurons are assumed to be plastic and it is this plasticity which allows them to learn meaningful, representational, firing patterns. Iconic Transfer: This key property relates to the source of information which controls the learning of the neurons. It will be seen that distal, sensory information is postulated to impose output patterns on neurons so that these may be learned and recalled in the absence of input. It is this transfer that creates inner perception in the conscious organism. Sensory Neurons: These are transducer neurons that transform energy from environmental input into the distal, sensory signals which control iconic transfer.

Corollary 1: The brain is a state machine.

The brain of a conscious organism is a state machine whose state variables are the outputs of neurons. This implies that a definition of consciousness be developed in terms of the elements of state machine theory.

Corollary 1 is a consequence of the intent in the fundamental postulate that the theory of artificial consciousness be based on state machine theory. State machines can model any system with inputs outputs, internal states and input-dependent links between such states. The states and their links form a state structure. Such machines can be probabilistic where links between states are defined as probabilities, they can have a finite or an infinite number of states. The fact that any conscious organism must have something called a brain with an attendant state structure is evidently true and not controversial. The key question is whether enough can be said about the nature of the state structure of organisms that are said to be conscious which distinguishes consciousness itself. This becomes the task for the corollaries which follow - to define the characteristics of state structure that are necessary for and specific to organisms that are said to be conscious.

Formalization of Corollary 1.

In any state machine, five items need to be defined:

i) The total input to the neural state machine is a vector i of input variables i1, i2 ...

The i1, i2 ..variables are the outputs of sensory neurons.

In living brains the number of such variables, being the number of neurons involved in the early layers of all sensory activity, is very large but finite. There is also some debate about whether it is important for these variables to be considered as binary (firing or not) or real (firing intensity per unit time). While it will be seen that this decision does not alter the course of the theory, it is assumed here that these variables are binary. This is done without loss of generality but with the gain that, using the methods of automata theory, it becomes possible to develop non-linear models.

Also, I is defined to be the set of all possible input vectors.

ii) The total output of the neural state machine is a vector z of output variables z1, z2 ...

The z1, z2 .. variables are the outputs of 'actuator' neurons.

Again the variables z1, z2.... are considered to be binary, and, in living brains, would be seen as the output parts of the brain which are responsible for muscular action.

Also, Z is said to be the set of all possible output vectors.

iii) The inner state of the neural state machine is defined as a vector q of variables q1, q2 ..

The q1, q2 ... variables are the outputs of 'inner' neurons. Again, variables q1, q2 ... are binary, and, in brains, would be the states of neurons neither involved in input sensing nor output generation.

Also, Q is said to be the set of all possible input vectors.

iv) The state dynamics of the neural state machine are determined by the equation

where q' is the "next" state, q is the current state and § a function, which in the case of a finite number of binary variables may be expressed as the mapping, where x is the Cartesian product. (In the general case this mapping is considered to be probabilistic in the sense that every pair (q,i) of Q x I maps into every element of Q with some probability.)

v) The output function of the neural state machine is determined by the equation

where w is a many-to-many mapping which in the general case is probabilistic.

In addition to the above group of five properties which are required in the definition of any state machine, the definition of a neural state machine contains the following key property which relates to the generalization of the neurons.

vi) The state dynamics and output functions of the neural state machine generalize in the sense that:

That is, there is an equivalence class of pairs (qj,ik) which includes (qa,ia) for which the next state qj is the same and the output za is the same. Note that the case for the stable state has been quoted here, although any state transition has similar equivalent state-input pairs. More is said of generalization in corollary 7 and appendix 1. The basic notion expressed in this corollary is expressed in fig.1.


Fig.1 A classical element of state structure. It is stressed that the transition is learned and that the state, input and output vectors are composed of neural state variables. ____________________________________________________________________________

The following three corollaries are stated together because their justifications are interleaved.

Corollary 2: Inner Neuron Partitioning

The inner neurons of a conscious organism are partitioned into at least three sets: Perceptual Inner Neurons: responsible for perception and perceptual memory; Auxiliary Inner Neurons: responsible for inner 'labelling' perceptual events. Functional Inner Neurons: responsible for 'life-support' functions - not involved in consciousness.

Corollary 3: Conscious and Unconscious States

Consciousness in a conscious organism resides directly in the perceptual inner neurons in two fundamental modes:
Perceptual: which is active during perception - when sensory neurons are active;
Mental: which is active even when sensory neurons are inactive. The activity of the inner perceptual neurons ranges over the same states in both these modes.

The same perceptual neurons can enter semi-conscious or unconscious states that are not related to perception.

Corollary 4: Perceptual Learning and Memory

Perception is a process of the input sensory neurons causing selected perceptual inner neurons to fire and others not. This firing pattern on inner neurons is the inner representation of the percept - that which is felt by the conscious organism. Learning is a process of adapting not only to the firing of the input neurons, but also to the firing patterns of the other perceptual inner neurons. Generalisation in the neurons (i.e. responding to patterns similar to the learnt ones) leads to representations of world states being self-sustained in the inner neurons and capable of being triggered by inputs similar to those learned originally.

Comment on corollaries 2, 3, 4.

i. All three corollaries stem from the statement in the fundamental postulate that a conscious organism is conscious through owning the sensation-causing firing patterns of its inner neurons.

ii. All three corollaries meet the requirement that an organism could not be said to be conscious unless sensations due to sensory input may be sustained in the absence of such sensory input, albeit in reduced detail. (The organism is conscious even with its eyes closed and other senses shut off).

iii. Allowing for unconscious function in a brain-like organism, corollary 2 indicates that perceptual states occur in a subset of inner neurons. That is, not all inner neurons store perceptual memories - some may be encode concepts such as duration, ordinality or even 'mood', while others just keep the organism "alive".

iv. Corollary 3 indicates that the fundamental postulate leaves open the possibility that not all the states of the perceptual inner neurons have direct sensory correlates. This allows the model to account for effects such as sleep or anaesthesia.


Fig.2 The classical state machine has its state vector partitioned into perceptual, auxiliary and functional variables. ____________________________________________________________________________

v. Corollary 4 suggests that the formalization of the fundamental postulate should account for the creation of perception-related states by reference to the learning properties of the neuron. This includes a formalization of the process of retrieval of inner perceptual states. Formalization of corollary 2

Let {q} denote the set of variables that make up vector q.
{q} is partitioned into three subsets:

Related to these partitions are the sets of all states on the variables:

Also Q = Qp x Qa x Qf

Fig. 2 shows what is intended physically by this partition.

Formalization of corollary 3

A particular perceptual input iw has a state correlate qw on the set of variables {qp} so that:

Also if iw is replaced by some "neutral" input i¿ , due to generalization (corollary 1 [vi]) (more will be said about the nature of "neutral" inputs in conjunction with corollary 10 relating to "will").

Let the set of all input-related states such as qw be Qi, a subset of Qp. The remainder of Qp contains states not related to perception: unconscious and semi-conscious states.

An assumption has been made in the above, and that is that

This effectively assumes that vector qp is independent of the rest of the state variables. This has been done to simplify the formal notation in this and subsequent corollaries and to retain a clear line of explanation. It is stressed that this assumption should not be made in general, as dependence between state variable sets provides a way of explaining links between perceptual states on one hand and auxiliary or functional states on the other. The general idea is illustrated in fig. 3.


Fig. 3. The upper example suggests that a meaningful transition may be sustained even in the absence of perceptual input. The lower example suggests that the same state variables as those involved in conscious perception are at work in a less controlled way when an unconscious sensory environment j¿ is present. ____________________________________________________________________________

Formalization of corollary 4

Learning is the process of first associating an input iw and an arbitrary state q¿ to form the following element of the forward network function §,

The key factor is that there is µ, a fixed sampling mapping which transfers some of iw into qw, causing qw to be defined by iw: To create the stable representation (attractor) for iw , § is augmented by: (This is the "iconic" training methodology fully described elsewhere [6] ).

Further, the generalization of the neurons as formalised in corollary 1, ensures that the requirement of corollary 3:

is satisfied. Figure 4 illustrates this as a state structure.


Fig. 4. State qw is formed to retain the spatial characteristics of iw which may be retrieved even in the mental mode. ____________________________________________________________________________

Corollary 5: Prediction

Relationships between world states are mirrored in the state structure of the conscious organism enabling the organism to predict events.

Prediction is one of the key functions of consciousness. An organism that cannot predict would have a seriously hampered consciousness. It can be shown formally that prediction follows from a deeper look at the learning mechanism of corollary 4.


Fig. 5. Planning is seen to be effective for mental input i¿, ____________________________________________________________________________

Formalization of corollary 5

Say that ix follows iw as a result of world state changes. Say that the organism is in state qw in response to iw. If the input changes to ix and iconic learning is taking place, the following element of § will be added:

In the mental mode, the following two transitions become equally probable: This means that, in time, state qw will lead to qx in the mental mode, completing the prediction as shown in Figure 5.


Fig.6. Internal and external loops which show how the effect of two actions is learned. ____________________________________________________________________________


Fig.7. Internal and external loops which show how the effect of two actions is learned. ____________________________________________________________________________

Corollary 6: The Awareness of Self

As a result of iconic learning and feedback between physical output and the senses, the internal state structure of a conscious organism carries a representation of its own output and the effect that such an output can have on world states. This includes a representation of what can and cannot be achieved by the organism itself.

Awareness of self is the ability to distinguish between changes in world states that are caused by the organism's own actions and those that occur in a way that is not controlled by the organism. Here it is argued that this ability follows from the prediction corollary and implies that the organism stores the knowledge.

Formalization of Corollary 6

The objective of this formalization is to involve the output of the organism as part of the input.

Let z¿ be a special "no output" condition and let all other output actions perceivable by the input be: z1, z2 .... zj An input to the neural state machine therefore contains at least two components:

[p = {r¥s} should be interpreted as "p is made up of two fields r and s"]

(note that if the output is z¿ then iz¿ = iw , and that is the situation assumed in earlier corollaries)

As iconic learning defines qw = µ( iw ) , this can be extended to

Hence iconic learning leads to parts of states such as sj which are internal representations of the organism's own actions.

Now suppose that the world is in some state ia and that this has been learned with z¿, producing therefore a direct iconic representation qa = µ( ia ), suppose that action z1 changes the world state to i1 where it remains even if the action ceases (i.e. the output reverts to z¿) . As in the prediction corollary, 5, the system will learn the following linked internal representations.

Hence iconic learning leads to representations of the way in which the organism's own actions achieve changes in the world state which is an existence proof for a representation of the awareness of self.

Figure 6 shows how the internal and external loops create representations of the way in which action za changes the world state to ia and that this is represented as qa. A similar representation is shown for the consequence of action zb. Figure 7 shows how, in the mental mode, this "awareness of self" may be retrieved as a probabilistic event.

Corollary 7: Representation of Meaning.

When sensory events occur simultaneously or in close time proximity in different sensory modalities, iconic learning and generalization of the neural state machine ensures that one can be recalled from the other.

When it is said that a conscious organism knows the meaning of input (e.g. words or events) this refers to the creation of an internal representation that is related to that input. It could be the word related to an event or an entire internal scenario related to a word. In this theory, the word "knows" translates to the retrieval of complete iconic internal representations from partial inputs. Association of this kind is a basic property of an iconically trained neural state machine and is formally stated below. However, this is also the area in which much of the controversy of whether a machine can or cannot "know" the meaning of words and phrases arises [7]. While linguistic output and the understanding of language are discussed in the corollaries that follow, it may be stated here that internal representations have precisely the property of subjectivity that Searle [3] has advocated. This results from such representations being iconic and dynamic "images" of world states which include the representation of world behaviour due to the actions of the organism itself. This is not the case with the preprogrammed symbolic representations of conventional Artificial Intelligence where the symbols are only arbitrarily related to sensory input.

Formalization of Corollary 7

The corollary has been worded so as to embrace two major forms of association of meaning: i) spatial and ii) temporal. These are treated separately.

i. Spatial association.
Here two input events, say iv and iu , (e.g. vision and utterance) occur together so as to be the components of an overall input iw. Say that, using the notation of corollary 6:

and for iconic learning


Fig. 8. Spatial association. ____________________________________________________________________________

It is shown in appendix 1 that there exists a generalization such that

where i¿ and q¿ are arbitrary values.

This leads to the stable states

This means that inner representation of the pair of associating events is recalled from the presence of one input event from the pair as illustrated in Figure 8.

ii. Temporal association.
In this case, input events such as iv and iu occur in sequence, and the retrieval of the second could be regarded to be the prediction made by the system as a result of the occurrence of the first. The formalization is therefore the same as that in the Prediction Corollary, 5.

Corollary 8: Learning Utterances.

The feedback loop responsible for the creation of "self" representations is also responsible for the creation of state representations of the basic utterances of the organism which are retrieved in response to the utterances of other, "adult" , organisms and may be used by the adult to teach the organism more complex utterances such as the words of a language.

A salient measure of the level of sophistication of an artificially conscious organism is the complexity of the means of communication that the organism has with others and the way that others may be responsible for teaching the learning organism to harness its abilities to communicate. In the spectrum of living organisms there appears to be a large gap in this sophistication between humans and animals which makes this and the next corollary appear to be particularly directed at human-like behaviour. In artificial systems it is possible to imagine a greater degree of continuity along this scale of sophistication than that which exists in nature. This is reflected in the formalization of this corollary.

The corollary involves the assumption that the organism is capable of a basic set of utterances, which, before any learning takes place, are generated arbitrarily. However, due to the output - input long-routed feedback, these, in the manner of corollary 6, form representations in the state space of the organism. If an "adult" wishes to key in to these representations it will have to imitate the utterances of the organism. This, among humans, is called baby talk.

The corollary extends to the way in which an adult, having keyed in to these representations, can use them to label events with words. In corollary 10 it will be shown that the organism can use such representations when it "wishes" to call for objects described by these words.

Formalization of Corollary 8

Concentrating on a specific set of the actions found in corollary 6, assume that the organism is capable of a finite set of elementary "utterances".

There is again an elementary utterance zu¿ which corresponds to no utterance or silence.

Say that the organism emits these arbitrarily and in the absence of other input, they are perceived at the input and learned as stable states as

However, as learning affects the output neurons too, a link is established in the output mapping w (see corollary 1): In summary, this means that an utterance zuj has an internal representation suj which is both retrieved by zuj through mapping § and sustained by it through mapping w. Hence a stable state has been established in both the internal feedback in the neural state machine and the long-routed feedback via output and input.

The implication of the above is that once state suj is entered, the organism cannot exit the state and will repeatedly and uncontrollably output zuj. This problem vanishes if it is assumed that all utterances except zu¿ are of short duration and that the normal sustained state is that of silence.

Now consider that another organism (which we call an 'adult') is capable of mimicking zuj as, say, zuj' .

If, s(zuj', zuj )> s( zuj', zuk ), k ­ j ( s is 'similarity' as defined in appendix I), then zuj' can 'induce' suj and the utterance zuj in the organism through the generalization property stated in appendix I.

Finally, a 'word' is defined as consisting of a sequence of utterances taken from the set of elementary utterances Zu. Say that the adult utters the word represented by some sequence

with the object in question in view at the time, causing (say) iv. As a combination of corollaries 5, 6 and 7 the transition between states may be learned, where qv is the state representation of iv linked spatially to sua and sub which are the state representations related to the utterance sequence zua' , zub' ... (i.e. the "meaning" label which is word-like rather than a single symbol as in corollary 7).

It may be seen that being given the stimulus iv, the entry into states {qv,sua } -> { qv,sub}... can give rise to the organism 'naming' the stimulus as zua , zub ... through a generalization of the output link zuj = w (suj ). This effect will be elaborated in corollary 10 and is illustrated in Figure 9.


Fig. 9 Utterance learning with im being motherÕs face and zaÕ baby talk for utterance za. Mental interpretation: organism expects mother if za za is produced. ____________________________________________________________________________

Note that if the same utterance is repeated (as in mah mah for example) this would bring into play auxiliary states (see corollary 2) and the above transition would be of the form

where a1 and a2 are the auxiliary state labels that distinguish between repeated utterances.

Corollary 9: Learning Language.

Language a result of the growth process of a societal repository from which it can be learned by a conscious organism, given the availability of knowledgeable 'instructors'.

It is not the intention here to give a full account of language learning by a neural state machine. It is, however, necessary to stress the three points that are central to this corollary. First, it accepts the concept (as accepted by many commentators on language and consciousness1,3,11) that language is the result of a process of growth of a repository to which new words and structures are being added by the users of the language and which resides collectively in the brains of such users. Second, it calls for the presence of an instructor. It has been argued6 that an organism that merely explores or observes an environment cannot learn the rules that govern phrase structured languages. This is overcome through the intermediacy of a teacher who selects progressively more complex examples of phrases from which the rules can be learned. The role of such a teacher is demonstrated in the formalization of this corollary.

The third point, while not explicitly expressed in the corollary, is central to the nurture-nature debate in language. It is the "poverty of stimulus" argument8. This suggests that deep language structure needs to be innate as, otherwise, more stimuli than appear to be available to developing children, would be required. What needs to be demonstrated formally is that adequate rules may be developed by a learning state machine from a reasonable number of examples.

The role of the instructor is more complex than that implied in corollary 8. This corollary relies on the instructor not only to provide word labels for objects and actions, but also to create the opportunity for the organism to build up the experience that underlies phrases and sentences. For example the phrase "a bag of beans" has to be learned as being not only grammatically correct with respect to "bag a beans of", but also semantically correct with respect to "a bean of bags". The training should also distinguish the use of the word "of" in "the sound of thunder" and "a bag of beans". The formalization of this corollary concentrates these phrases as a particular example, while the learning of other linguistic objects (such as anaphora9) are the subject of current studies.

Formalization of corollary 9.

It follows from corollaries 7 and 8 that state representations can, spatially, within a state, represent both the appearence and the meaning of objects. In order to learn the rule for accepting "a bag of beans" and rejecting "a bean of bags", the neural state machine has to learn the difference between a "container" and the "contained". This follows from an extension of corollaries 7 and 8, in the sense that the representation of an object as a state can carry a large list of attribute labels that represent the experience of the organism, either first hand or through manipulative experience. In fact, the state of a neural state machine can partition its state variables into "fields" which (as in classical databases) carry attributes such as "non-porous", "hard", "concave", "open at the top", "container". In contrast with classical databases, however, given the first four, the fifth could be retrieved without it being given through the generalization of the system.

It follows that in the time domain, state qa1 ( = µ(ia1 ), the input word for, say, "bag") is followed by input word (hence iconic state) which is followed by qb1 ( = µ(ib1 ), the input word for, say, "bag"), and this is repeated several times for

and if qa1, qa2, qa3 .... learn "container", qcr, in common in some specific field (corollary 7), and if qb1, qb2, qb3 .... learn "contained", qcd, in common in some specific field (corollary 7), then the sequence < {qaj¥qcr} > < {qbj¥qcd} > will be learned while any phrase of the form < {qqcd} > < {qqcr} > will be rejected.

Note that in the above, only one attribute has been appended, but this attribute could be inferred from the presence of others through the generalization of the net. It is this that circumvents "poverty of stimulus" problem through the process of the dependence and independence of state fields, which can be established with a relatively small number of examples.

Corollary 10: Will

The organism, in its mental mode, can enter state trajectories according to need, desire or in an arbitrary manner not related to need. This gives it its powers of acting in a seemingly free and purposeful manner.

The effect which is modelled by this corollary is that which would lead the organism to describe a sensation as "I want X". This corollary follows from the representation of "self" in corollary 6. Suppose that a state, in the mental mode, can lead to several subsequent states. There are only two ways in which one of these transitions can be favoured. One is a purely arbitrary transition, and the other is some "desire" which sets the internal representation into the appropriate part of the state. The "desire" can be induced from some external stimulus such as someone's statement "do you want a chocolate or a vanilla ice cream?" or from some internal stimulus such as hunger. Whatever the case, the result is that the automaton in its mental state (as in corollary 5) predicts the result associated with each of the states constrained by the stimulus. This may evoke emotional links (see corollary 12: emotion) such as greater pleasure associated with chocolate as opposed to vanilla or instinctive links (see corollary 11: instinct) that lead to crying to alleviate hunger (interpreted by an observer as "baby wants food").

In an organism that has learned to utter words (as in the discussion of corollary 8), an example of two actions as discussed above, could be uttering mah-mah or not with the result of bringing "mother" into view or not. Therefore the sense in which an organism could learn to express will by saying "I want X" is precisely a "field" generalization learned from an adult as in corollary 9. The interpretation of such a statement is that the organism, for an unspecified reason, has X in a specific mental field, and takes the learned action to control the environment in a way that in past experience brought X to the fore.

Formalization of Corollary 10

Assume that the neural state machine is in some state qw from which it has learned that its own actions za, zb, .... lead to world states ia, ib, .... (corollary 6) leading to mental representations

where qa is a representation of world state ia, sa a representation of action za and aa1¥aa2 ... are additional attributes of the state learned as in corollary 7. The learned transitions from qw are of the form

Similar learning takes place for other states such as (qsab1¥ ab2 ...).

Corollary 5 has explained that in the mental mode these states will be visited either in some arbitrary fashion, or (as in appendix 1) can be retrieved in response to the presence of a subset of attributes.

That is, given state qw and the absence of all other attributes the system, in the mental mode, can transit to any of the states (qsaa1¥ aa2 ...) or (qsab1¥ ab2 ...) ... etc. This would be described as the system "knowing" that it can do

The existence of a preferred attribute will lead to one transition becoming more probable than others through learning. That is, a state label such as aa1 can be the distinguishing label which causes a learned transition bias to (qsaa1¥ aa2 ...) in preference to others. For example, if pleasure is preferred to pain (as will be seen in corollary 11) or the presence of the mother's face is preferred to the absence due to a bound attribute such as the alleviation of hunger, this will determine the more


Fig. 10. Knowing that the world can be changed to ia or ib. ____________________________________________________________________________

likely mental transition and consequent action. Figure 10 illustrates the state structure which represnts knowledge which allows either a "free" choice or a choice according to need. The right hand side of the figure shows how the mental model leads to predicted action.

Corollary 11: Instinct

To enhance survival, an organism needs a substrate of output actions that are instinctively l inked to inputs. These form anchors for the growth of state structure which develops as a result of learning.

There is no doubt that living organisms are born with instinctive reactions to some inputs. This is particularly true of animals who need these instincts for early survival. Typical of such links is the suckling reaction in response to a feeling of hunger and the presence of the mother's breast. In an artificially conscious organism, such reactions are to be determined by the designer. The thrust of this corollary is that such instinctive reactions influence the development of learned state structure in the neural state machine.

Formalization of Corollary 11.

An instinctive reaction is of the form

The way in which this affects state structure, soon after "birth" is (as in corollary 6) through the iconic creation of a state which becomes linked to the action za through learning as This means that a representation of such instinctive links makes an early entry into the system's state structure. This influences the development of semantic learning, utterance and language acquisition as described in corollaries 7, 8 and 9, above. Figure 11 shows the za = l(ia) effect.


Fig. 11. A "instinctive" output reaction to a world input happens without a change of 'mental' state. ____________________________________________________________________________

Corollary 12: Emotion.

Auxiliary state variables encode emotions which develop from built-in instincts.

The presence of auxiliary neurons has been outlined in corollary 2. In living organisms instinct can account for basic sensations such as fear and pleasure. As such sensations are not representative of world events they must be encoded in auxiliary neurons, primarily in an instinctive way as defined in corollary 11. However, as learning progresses, these encodings become associated with mental representations of world events. This process is responsible for both the spreading of such codes to states that are not only reactions to world events but are related to such events through the process of prediction. Subtle, mood-like emotions can be modelled in this way. Also it may be possible to model pathological cases where the emotion labels have spread to inappropriate mental states. This is a vast topic which requires a great deal more attention than is provided here. The formalization below merely indicates a way in which theory could be developed in this area.

In artificial systems, in much the same way as instincts, useful, basic emotions would have to be determined by the designer. Formalization of Corollary 12.

A primary emotion is instinctively linked to some input as

where af is a state of auxiliary neurons.

For example, af could be "fear" and if the image of a looming object.


Fig. 12. Auxiliary emotional states representing pleasure and pain. ____________________________________________________________________________

As in the case of instincts, iconic learning will lead to internal representations of the world events to be labelled by the emotional states, for example, if

iconic training will lead to internal representation of the kind which can be accessed in both perceptual and mental modes. Figure 12 shows the linking of two different auxiliary messages with two world states. From this point onwards the prediction and self awareness processes formalized in earlier corollaries come into play, associating the auxiliary states with state trajectories (thoughts) related to world events. The mixtures of auxiliary states and mental representations of world states leads to subtler forms of emotions. A simple example is the transfer of the pleasure of the alleviation of hunger to the representation of the mother's face. It is this property that, through the will corollary (10) endows the organism with "pleasure seeking" or "pain avoidance" behaviour.

Consider now the example of a pathological case (say, depression) which may be linked to the instinctive emotion of "unhappiness", say, au. It is possible for the au sensation to become linked to a large number mental states on the perceptual neurons. Neural theory can show that under some conditions, the au code can become self-sustaining on the auxiliary variables and be present even for unrelated states on the perceptual neurons. An observer would ascribe "chronic depression" to this organism. The theraputic act of a counsellor could be described as trying to recreate the links of au to appropriate mental states and liberate the inappropriate ones from it.

2 Technical critique

In this section and the one that follows the theory presented above will be given the shorthand title of ACT (Artificial Consciousness Theory). ACT is by no means rigorously complete. Its aim is to establish, through what could be described as "ramshackle" formality, that there exists a theoretical framework that is appropriate to a description of consciousness. Stress has been laid on the word "artificial" to keep in mind the constructional nature of the exercise: given an engineered artefact, what properties must it have to capture consciousness? Particular care has been taken not to fall into the trap of saying that only a biological brain can have this property as this endows the brain with mystical properties that set it apart from the physical world and which require analytical methods that have as yet not been discovered and, indeed, may never be discovered. The diametrically opposite point of view has been taken - the brain is merely one of a large class of conscious mechanisms, some of which may be manufactured. These operate according to a set of principles that may be described using existing analytical methods. However, it is the existence of the biological brain that inspires and informs the content of ACT and ACT itself can be applied to a detailed analysis of what is known about biological brains. ACT also asks a more general question: say that the universe of mechanisms could be divided between those that can be said to possess consciousness and those that do not, what is the difference between the two and what is a canonical model for the class of mechanisms that have consciousness?

It is the fundamental postulate that contains the belief the consequences of which lead to the required distinction. That is, consciousness is an inward sensation which can be attributed to the activity of a subset of constituent components called neurons. In living organisms the processing machine made up of the totality of neurons is called the brain. While this assertion is shared by many contemporary views of consciousness (see 3, below) the consequences of such a belief have not been previously analysed, and this is the objective of the 12 corollaries developed in this paper. The language of probabilistic automata and a general ,physically realisable model, the neural state machine, have been used to show in the corollaries that attributes of consciousness can be captured with a constructional methodology. Clearly, the list of corollaries need not be limited to 12. In fact it is important to keep the list open to encourage the further development and refinement of the theory. It is evident that the corollaries are ordered, the earlier ones establishing properties on which the later ones can build.

A question that remains to be answered here is whether the approach in this paper explains the richness of conscious experience which we as humans enjoy in the course of introspection. It is difficult to be quantitative about this, but it can be shown that the artificial model discussed here would certainly provide both states and a state structure of remarkable richness due to the fact that the number of states of an automaton increases exponentially with the number of state variables (neurons). Assuming that states can be visited at the rate at which live neurons fire (about a 100 times a second) over a period of 100 years, the automaton would be able to change state approximately 4x1011 times. Now, as an automaton with n state variables has 2n states, and 4x1011 is approximately 236, it would take over a hundred years for a modest automaton with 36 neurons to visit all its states. Of course the nature of neural state machines in ACT is such that only a small subset of all states is used to represent world events. Despite this, the key point is that ACT can represent mental states in a very rich way which should be contrasted with models in traditional artificial intelligence that are highly parsimonious. Also, the richness allows for complex functions to be built up from absorbing a large number of events, which provides an explanation of how it is that a mass of brain cells can learn to behave in ways that are modelled by the parsimonious, symbolic representations created by a programmer's efforts in traditional artificial intelligence.

3 Conclusion: philosophical critique

ACT appears to clash specifically with Penrose's contention [2] that consciousness is such an important phenomenon that it is hard for anyone to believe that it is "something just accidentally conjured up by computation". While ACT is undoubtedly couched in terms of a computational framework, there is no appeal to an accidental emergence of consciousness. Indeed, the main aim of the theory is to show that the complex mixture of properties normally attributed to a conscious organism are the properties associated with some computing structures and may be described through appropriate formalisms. Part of the problem lies with the word "computer" as this is taken by Penrose [4] to mean "machine that operates according to rules designed by a programmer". But that is only one interpretation of computation. Here the alternative of "neural computation" forms the basis of a theory in which the programmer merely plays the role of creating machinery that is capable of absorbing and representing its environment. The programmer does not feature in the theory itself which is about the state structures that the computational organism creates for itself or through the aid of other conscious organisms. Therefore while it is possible to agree that consciousness cannot be captured by a programmer's recipe (algorithm), the door should at least be kept open for computational models of consciousness based on systems that are capable of building up their own processing structures.

At the other end of the spectrum of belief in computational models, Dennett's explanation of consciousness1 relies very heavily on elements of computer engineering. The theory is expressed in terms of consciousness being the product of a multi-agent program running on a virtual machine where the autonomous "bossless" agents are continually reassessing sensory information ("revising drafts"). In comparison with ACT, this outlook appears strongly metaphorical. All the concepts are metaphors drawn from computer engineering, but as metaphors they do not necessarily constitute a theory. While the Cartesian ghost in the machine has been expunged, the ghost of the programmer is still there, and this does little to explain how the machine components come into being and do what they do. This is left to a briefly stated appeal to evolution. There is agreement between Dennett's description and ACT in what regards language being held in an evolving societal repository - an idea borrowed from Dawkins [10].

Evolutionary models of consciousness such as those of Edelman11 abound in the literature. The mechanics of the brain are seen as being the result of a very finely tuned evolutionary process which requires an accelerated evolution of neural structures during the course of life. This contrasts with ACT in the sense that learning in a neural state automaton does not demand the physical changes associated with lifetime evolution. The plasticity and generalization of the neurons is sufficient. An evolutionary standpoint is also part of the work of Humphrey [12], who sees conscious representations as evolving from the internalization of instinctive reactions to the environment. This seems an unlikely process as such internalizations have subsequently to be externalized in order to develop a sense of self and to use language as indicated in corollaries 6-12 of ACT.

The philosophical standpoint which does not clash with ACT is that of Searle [3] as, at least, it leaves open the possibility of finding the causal nature of the brain (i.e. causing consciousness through neural activity) and the subjective nature of consciousness (intentional relationship with the environment) in materials other than biological brains as we know them. While he goes on to believe that it is most unlikely that such alternative chemistries could be found, his philosophy does not exclude a theoretical analysis of alternatives as being of relevance to an understanding of biological consciousness. ACT is precisely such an analysis. It is based on an abstraction of what is understood to be the function of the brain as it attempts to capture both the causal (fundamental postulate) and subjective (corollaries 6-12) character of consciousness. Even if some artificial system built on ACT principles were to leave skeptics unsatisfied about it "actually being conscious", it is likely that this judgement would be based on the belief that only biological systems could be conscious. This would not reflect on the appropriateness or otherwise of ACT which addresses what it is to be conscious whether this is said of organisms that are biological or not.

Contemporary debate on consciousness has a much wider participation than that of the authors quoted so far. For example, Nagel's suggestion that it is necessary to say what it is like to be a particular conscious organism [13], can, in ACT be expressed in terms of a taxonomy of state structures (i.e. how does the state structure of a bat differ from that of a human?). Also the fundamental postulate in ACT contradicts McGinn's contention [14] that the link between introspective sensation and the function of the brain cannot be understood as a scientific object. It has been shown that the alternative view leads to a comprehensive theory which should be evaluated against the pessimistic result McGinn's proposal that consciousness cannot be understood through logical enquiry. ACT creates an opprotunity for a much longer exposition that includes the views and nuances of the many who have expressed opinions on the nature of consciousness. All that has been done here is to draw attention to the fact that this can be done in a principled way.

Appendix I: Generalization

Corollary 7 has asked for an existence proof for a generalization which causes an input pattern where one part is the same as a training input and the others are arbitrary to retrieve the appropriate response.

The existence proof of such a generalization makes use of a similarity measure s(a,b) that can be used to compare any two patterns a and b. This is a value which is at a maximum when the two patterns are identical and a minimum when the patterns are the inverse of one another. The measure could be expected to be half way between maximum and minimum for two arbitrarily chosen patterns. Also when pattern a is made of patterns k and l, i.e. a=(k¥l) and b=(m¥n) then s(a¥b) = s(k¥m) + s(l¥n). In neural technology, the negative of Hamming distance - the difference between two patterns measured in number of bits - is used as such a measure, and has the required properties, which indicates an existence of the assumed measure. Hamming distance also seems a good measure to describe similarity in living neural systems. Assertion 1: Ideal generalization is expected to yield qw in response to neural state machine input/state pair (ix , qx) if (iw, qw) is the most similar pattern (among a training set of such pairs) to (ix, qx) and qw = § (iw, qw) .

Referring to the situation in corollary 7 where the input is a mixed pattern of the form (iiu), contradiction of assertion 1 would require that there be some (iiy, qz) that is more similar to (iiu, qw) than is (ii¿, q¿). Remembering that qw = (qv, qu)this would lead to the erroneous retrieval of qz in place of the correct retrieval qw. But this would require that

As it can be expected that s(iy ,i¿) » s(qz ,q¿) » s(iu ,i¿) » s(qw ,q¿) ( » is used to indicate "roughly the same as"). The above inequality implies that s(ix,iv) > s(iv,iv) which cannot be as s(iv,iv) is maximum. It is noted that this is an expectation and not a guarantee, based on the expectation of the similarity of two arbitrarily chosen patterns.


1. Dennett, D. C. Consciousness Explained (Allan Lane/Penguin, London, 1991).
2. Penrose, R. The Emperor's New Mind (Oxford U. Press, 1989)
3. Searle, J.R. The rediscovery of the mind. (MIT Press, Boston,1992).
4. Penrose, R. The Shadows of the Mind (Oxford U. Press, 1994)
5. Kelly, G. A. The Psychology of Personal Constructs (Norton, New York, 1955)
Aleksander, I. and Morton, H.B.Neurons and Symbols: the Stuff that Mind is Made of. (Chapman and Hall, London, 1993).
7. Searle, J.R. Behavioural and Brain Sciences 3, 417-424, 1980.
8. Chomsky, N. Language and Mind (Harcourt Brace Jovanovich, 1972)
9. Parfitt, S.H. Proc. ICANN94, 188-194 (1994)
10 Dawkins, R. The Selfish Gene (Oxford University Press, Oxford, 1976)
11 Edelman, G. M. Bright Air, Brilliant Fire. (Allen Lane, London, 1992)
12 Humphrey, N. A History of Mind (Vintage, London, 1993)
13 Nagel, T The View from Nowhere (Oxford University Press, Oxford, 1982)
14 McGinn, C. The Problem of Consciousness (Basil Blackwell, Oxford, 1991)
15 Aleksander I., Towards a Neural Model of Consciousness, Proc ICANN 94, Springer (1994)