Interestingly, many of the problems I have just sketched seem to arise again if we do not look upon ourselves as phenomenal subjects with an inner perspective, but as information processing objects, namely as representational systems with a long biological history. For this reason, the hypothesis suggests itself that the presence and striking holism of phenomenal reality, which we have so far only noticed from the subjective perspective, would no longer have to be explained in accordance with classical philosophical models from above (e.g. by a transcendental subject), if we had a good bottom-up alternative. Possibly what is needed is a generalizable scientific theory of the capacity called 'feature binding' in the terminology of brain research: The fusion of different properties perceived by the system into a holistic internal structure. Such a binding of properties is, for instance, necessary in enabling us to see objects as objects. In the sensory modality of visual awareness for example, such properties could be edges, movements, surfaces or colours. As we know, these features are represented by spatially separated groups of neurons in the brain. In subjective space however, they appear as a phenomenal Holon - for instance as a book in one's hand. For this reason, the question arises of how the different flows of information can be integrated to form a unified data-structure.
This binding problem is one of the central problems in brain research. Its logical structure resembles classical debates in philosophy and psychology (e.g. in the theory of objects or in Gestalt-theory). Models of neural networks frequently represent the local features of a perceived object through activation states in property spaces, which are physically realized in the brain by scattered, spatially non-adjacent areas. For this reason, simple neighbourhood-interactions between single nerve cells cannot help the system to integrate the different properties of the object into a representational whole. Not only our subjective consciousness, but also the overall representational state of a connectionist system can be seen as an ascending hierarchy of such wholes: of elementary properties - the counterparts of subjective colours, bodily sensations or sound experiences - of objects, scenes, situations, contexts, a model of the self and a model of the world, in which the former is located. If we thus look at our brain from the outside and describe it as an information processing system which generates internal representations of the world and of itself, the binding problem poses itself on a multitude of levels.[18]
The brain responds to stimulation by a coherent object - for instance in visual perception of an external object - with a multitude of spatially distributed activation patterns. In order to differentiate between different objects or to detach representational Gestalten from a background, the system has to achieve an integration of these spatially distributed events into a single, ordered and unified pattern of activity, without causing a 'superposition-catastrophe'. There is another theoretical problem, closely connected with the binding problem, the importance of which can hardly be overestimated for any theory of phenomenal consciousness: the superposition problem. In order for more than one bound pattern of activity to coexist in a system, no interferences or wrong connections of properties must occur. The different patterns must not delete each other. A mechanism which solves this problem is, for instance, required in order to separate a figure from the background in a visual picture and to segregate it from different figures. Only when both problems, the superposition problem and the binding problem, have been solved can one imagine how a holistic representational state, a unified object (which is then embedded in further and more extensive states of the same type) emerges from the activity of many spatially distributed feature detectors. Let us call such a state a 'representational Holon'. Additionally, this 'representational Holon' has to be functionally active, i.e. it must be able to play a causal role as a coherent whole emerging from properties represented through singular events, for instance in producing co-ordinated behaviour. In order to generate such stable representational states it is therefore necessary for the activity of large populations of neurons to be co-ordinated in their parallel and highly specific activity. This is the neurobiological aspect of the binding problem and the superposition problem.
A number of different theoretical models for the solution of this problem have already been proposed, such as Barlowian pontifical neurons[19] and Donald Hebb's early assembly concept. At present, one of the most promising approaches is the 'correlation model'. According to this model, one assumes that coherence - for instance of a perceived object - is temporally coded, i.e. that the respective feature-sensitive cells represent this coherence through a precise temporal correlation.
Contrary to the classical model of assembly formation by coactivation, such a temporal coding would actually allow us to solve the binding problem, because the synchronization of neural impulses provides an additional variable for the structuring of neural patterns of activity. The temporal correlation in this case would exactly be the [. . .] selective 'label', which unequivocally specifies which subset of the activated neurons is bound together into an assembly. The overall pattern of active cells in the visual system would in this way obtain an inner structure, functionally meaningful for other regions of the brain, which it lacks in the Hebb-model. (Engel 1994: 13. English translation TM.)
The central assumption is that the coherence of perceptual objects is achieved by a synchronization of the firing rate of those cells which are sensitive to the respective features. According to this model, the temporal correlation within synchronously firing cell-assemblies would be the 'glue' through which these events in the system are 'bound'. As such a labelled and integrated whole, they can then enter further processing. Furthemore, such an expansion of the Hebb model by introducing a temporal dimension of coding also provides a mechanism for effecting important functions like figure/ground separation and differentiating between objects: namely by desynchronizing different cell assemblies. Since such a mechanism of dynamic feature-binding through a transient synchronization of spatially distributed cellular responses could be very fast, it would also increase the flexibility and the dynamics of the overall system in an economical manner. Before I move on, I would like to hint at why such a theoretical approach is also interesting for a philosophical theory of mind, e.g. under the aspect of ontological parsimony. If we have a conceptually consistent and empirically plausible model of feature binding (i.e. of the formation of representational objects as a form of self organization), this supplies us with the first ingredients for a naturalist theory of consciousness - that is to say, for a bottom-up explanation. Valerie Gray Hardcastle has aptly called this the possibility of giving a neo-Humean answer to an old Kantian claim.[20]
First signs pointing in this direction have now appeared, and they are of great interest for a philosophical theory of consciousness. For instance, Wolf Singer and his colleagues at the Max-Planck-Institut fur Hirnforschung in Frankfurt have discovered that spatially distant neurons in the brains of cats which react to stimuli originating from the same visually presented object, begin to oscillate synchronously with a frequency of 30 to 80 Hertz.[21] The suggestion that the binding of visual properties could be achieved through very short synchronizations of distributed patterns of activation had already been made by Christoph von der Malsburg in 1981 [22] These new discoveries show that synchronous neural oscillations of the field potential with a frequency of around 40 Hz establish themselves for very short periods (less than half a second). However, sometimes the role of these oscillations has been misinterpreted. In our context, the fact of synchronization itself is what is most important: In some cases, the oscillations can be understood as boundary conditions of synchronization processes, by which the integrative functions of the cortex may possibly be achieved. It is interesting to note that in the course of their self-organization these processes tend to respond to the classical Gestalt criteria such as neighbourhood, similarity, continuity of motion and so forth, and further that in some types of experimental set-up they take roughly the same time as that in which a person's attention jumps from one object to another. Researchers are also investigating whether such a high-resolution temporal code could be used to represent relations in non-propositional formats and to carry out higher-order object-constructions as well. This would be especially interesting for a general theory of mental representation, as it would not only clarify how the human mind forms objects out of distributed sets of properties, but also how it can episodically embed these in each other.
It is easy to see that the output of such coherently active cell assemblies could in turn be used as input to other 'coherence detecting' nets at higher levels, and those could, in turn, self-organize their connectivity as a function of the spatially and temporally structured input provided by the preceding processing levels. Iteration of such segmentation and regrouping operations could then allow for the generation of non-isomorphic, abstract representations of complex shapes and patterns. (Singer 1989a: 26.)
One could imagine higher-order embedding relations as a sequence of anatomic modules in series, but also as a dynamic hierarchy of activation patterns. Thus the theoretical principle of temporal coding might not only solve the binding problem and the superposition problem, but could also be operating on a multitude of representational levels. For this reason, a general mechanism of integration of this type would be of great interest in searching for an explanation of the holistic character of the overall representational state.
We rather have to ask, who is the subject of perception, and how is the unity of perception established in the brain? It is an ineradicable misconception that the unity of perception has to be established in a separate center, which in addition is often imagined as being of structureless unity itself. This mental archetype leads to infinite regress and to absurdity. Instead, the unity of mind has to be seen as an organic equilibrium among a great multitude of elements. The mental symbols both send and receive at the same time. Signals sent by one sub-symbol are deciphered by other sub-symbols, and the sending symbol can in turn establish itself, momentarily, if it responds to the messages and questions sent by others. In the state of unity, each subsymbol encodes in its own terms the situation described by others. This unity is not reached by leaving out detail but by uniting all detail with the help of relations. (von der Malsburg 1986:175.)
This remark of Christoph von der Malsburg illustrates impressively how the correlation theory of brain function could provide us with the conceptual means to reach a satisfactory naturalist answer both to the homunculus problem and the classical question concerning the unity of consciousness.
The technical details are less interesting for the philosophical issue than the new image of the brain resulting from them: Our brain is a system which 'escapes into time' above a certain level of representational content.[23] Let me try to illustrate this principle by introducing a new metaphor, although I will have to surrender it at a later stage of the discussion. This new metaphor is the time-window metaphor. My claim is that we are systems which generate meta-representational knowledge about some of their own states by opening time-windows of different sizes, through which they can look at the way in which their own autonomous activity is being modulated and structured by the information flow from the sense organs. Time-windows are neural integration windows: They provide a precise time scale for representational binding mechanisms.
The opening of time-windows is something which can happen from the outside. If neuroscientists observe the activity of the brain, they can calculate the average of certain magnitudes over a period of time in describing measurement data, or they can add up the results of a number of stimulus cycles (for instance when counting the frequency of action potentials in successive windows of 100 ms or when calculating correlograms in windows of 1 to 3 seconds). In doing so, they generate an abstract object of their representation, picking out a particular property of the system because they want to investigate it further. This property is intersubjectively accessible, is normally given in the form of a linguistic or propositional (e.g. mathematical) description, and emerges from the choice of a specific time-frame in the design of the experiment and in the interpretation of the data gained. The design of the experiment as well as the interpretation can be misleading relative to the epistemic goal in question. The scientific community could discover this at any time.
Importantly, it is possible that the brain itself opens time-windows on its own activity, for instance through the mechanism of temporal coding which I have just sketched. Through this process, a representation of spatially distributed micro-events within the brain is generated 'from the inside'. The 'inwardness' of this representation consists in the fact that the system in question automatically distances itself from its own physical 'processuality', because object formation filters the event-character of the underlying causal mechanisms out. To put it differently: Certain configurations of stimuli initiate dynamic processes of self-organization within the system. These processes converge towards higher-order states [24] which are 'labelled' as wholes through the synchronicity of the neural responses. They represent the coherence of the stimuli through the formation of a transient object, i.e. by way of a new and likewise coherent state within the system. Thereby the event-character of the underlying process of self-organization is filtered out and becomes unrecognizable for the system itself. Through synchronization, through the generation of temporal correlations, many spatially scattered neural responses (events) can be integrated into a higher-order whole (an object) appearing in the time-window. Many events become one object and by virtue of this, something like the 'surface of the inside' emerges.
This form of the opening of time-windows also singles out certain properties of the data flow within the system and bundles them together by binding them into an object. This object can then be investigated by further higher-order forms of information processing and representation. This temporally coded function of representation by feature binding is realized through a concrete state within the system itself, such as an activation pattern which is bound using synchronization processes. Again, a certain time frame is utilized, an object of representation emerges, and this object is, in principle, intersubjectively accessible. In this case, however, the 'design of the experiment' as well as the 'interpretation of data' are internal states of the cognizing system itself. Of course it is possible, that the underlying processes are subjectively inaccessible to that system itself because of peculiarities of its functional architecture.
Maybe it is possible to clarify further the concept of object-formation by pointing to its epistemological aspects: The high internal correlative strength or coherence of a perceived set of properties is a fact, normally in the environment of the system. This fact however, is not represented through a propositional format or the activation of a sentence-like structure, but through the generation of a holistic object, a representational Holon. It is for this reason, that in many cases the system can no longer distinguish between form and content - there is no syntax and no semantics. Thus, it seems that with the generation of representational objects of this type, an interesting non-conceptual form of 'abstraction by integration' is simultaneously achieved. For the system itself, however, the result of this abstraction appears as a concrete object.
A time-window is therefore an integration window in which the brain unifies different internally represented features into a whole by generating a temporal correlation. But the 'embedding' of the content of different time-windows in each other must be thought of in terms of a liquid architecture, in which plasticity and flexible overall dynamics can coexist with homogeneity and stability of form. Different types of time-windows could realize different 'grains', different resolutions by means of which various entities can be represented. However, as all of this is sub-symbolic information processing, time-windows must not be conceived of as rigid basic elements: we have to understand them as plastic parts of a dynamic binding mechanism, i.e. as variable conditions of representation, which, through this very variability, can also be context-sensitive and which are, at most, characterized by a weak compositionality. For this reason, the higher-order embedding of mental content must presuppose a dynamic evolution of the overall state, in the course of which bound patterns of activity can be superimposed without loss of any relevant information. Time windows could thus be plastic mechanisms. A - more or less strong - synchronization would then be the - more or less sticky - glue expressing the internal correlation-strength of the sets of properties represented by means of a synchronization gradient. This synchronization gradient lifts the representational object - more or less clearly - out of the background of activity which generated it. [25]
Now, let us pursue further the considerations sketched above by generalizing the idea. One can analyse the construction of a 'representational Holon' as precisely that case in which the system frees a distributed set of properties of the temporal difference relations that hold between its elements. This synchronization of single events within the system then results in the formation of a new and integrated form of content (i.e. in the generation of a representational whole by discrete mechanisms of temporal coding) on the respective higher levels of representation. Since the single events which function as feature detectors are endowed with a common temporal Gestalt by the system, the set of properties partially loses its internal temporal structure and is transformed into a holistic representational object: Synchronicity generates wholeness.
We are now able to extract a conceptual principle from these empirical considerations, one which could be highly interesting for the philosophy of mind and the naturalization of phenomenal consciousness. I will call this The Principle of the Formation of Representational Wholes (PFRW):
PFRW: Certain natural representation systems are able to bind internal, spatially distributed individual events which function as feature detectors for them into a representational whole, by coding the perceptive relations between them through processes of synchronization.
I will not concern myself here with a precise demarcation of the class of systems in question. [26] However, I will assume that human beings in non-pathological waking states belong to this class. At this point, after having looked at the problem of phenomenal and representational wholes from the first-person perspective as well as from the third-person perspective, I will return to conceptual analysis, offering philosophical and not empirical speculations.
First, I will use the cavalier audacity of a philosopher to make a very strong general assumption: The form of temporal coding postulated by the correlation theory is the general integrational mechanism by means of which - at least in systems of our own type - all kinds of representational wholeness are generated. In other words: I will tentatively assume that the mechanism I have just described (undoubtedly in a far too short and imprecise manner), as the process by which the system is able to open time-windows of different sizes on its own internal activity, is a general principle, operative on all levels of binding. My aim is not, however, to generate meaningless empirical pseudo-hypotheses. Rather, the goal is to investigate the heuristic potential of a theoretical model which has been deliberately generalized. For this purpose I shall make the following assumptions:
Al Temporal coding synchronizes the activity of spatially distributed feature detectors and, in doing so, permits the homogeneous presentation of one property in one sensory modality. ('Generation of elementary qualitative units within separate property spaces.')
A2 I will here combine five assumptions which are relevant to the representational construction of a complex external world and to the regulation of behaviour in interaction with this external world:
A3 The third assumption regards the representational construction of an integrated inner world. Temporal coding enables the system to generate a holistic self-model and to embed this model into complex situations. ('Centering of the overall representational state'; formation of the 'first-person perspective.')
A4 Temporal coding enables the system to bind the overall representational state which is already centred by a self-model into one global structure, i.e. into a highest-order coherent representational whole. (Highest-Order-Binding; 'Formation of a holistic reality model'.)
At present, these assumptions are at best weakly supported by empirical evidence. They cannot count as serious empirical speculations, and they are not intended as such. However, the empirical situation already justifies a rational philosophical speculation: It seems reasonable to investigate a new outlook on to familiar problems by generalizing an empirically plausible principle. It is instantly obvious that assumptions Al and A4 may be relevant to the elementary homogeneity of qualia and the global holism of the space of conscious experience. These questions will remain the focus of my considerations. However, because I have generalized the problem of the holistic character of phenomenal content by introducing the concept of a 'phenomenal Holon', the conceptual speculations too will now have to be generalized with regard to the respective levels of description.
First of all we therefore have to formulate a generalized new version of PFRW in correspondence with the assumptions Al to A4. This new version has to meet two requirements: First, it must abstract from our original example, viz. object formation within a single sensory modality. Secondly, it must allocate a causal role to the representational wholes which have been generated. Phenomena of synchronization are only interesting if it can be demonstrated that they play a separate functional role in the genesis of behaviour and in subjective experience. [27 Let us call this the Generalized Principle of the Formation of Representational Wholes (GPFRW):
CPFRW: Naturally emerged representation systems of a certain type are able to bind a subset of internal, spatially and temporally distributed individual events or representational wholes, which are already active, into a higher-order representational whole by coding the perceptive or embedding relations between the elements of the set through processes of synchronization. Sets of single events bound through a common time structure can in this way play a separate causal role in the system, i.e. they can sometimes be regarded as higher-order functional properties or transient functional modules. In this case they are functionally active representational wholes.
According to this principle, the binding mechanism which on all representational levels joins together single elements and episodically transforms them into higher-order wholes, consists in the generation of an identical time structure. If the representational 'levels', of which I have just spoken so carelessly, really are separate stages of the overall dynamics, then it should be possible to describe them by means of discrete classes of neural algorithms. Thus we have to investigate the time constants of the neural classes of algorithms in question. These time constants will be the conceptual essence of any abstract analysis of what 'holism' means on the representational level. A further question: In what cases are the functional states under consideration realized by clearly delimited anatomical modules on the level of the brain at all and in what cases are they only realized as transient computational modules?
At this stage of our considerations the time-window metaphor, which I have used to illustrate the principle of temporal coding, has to be abandoned. It could easily generate confusions since the opening of windows is an intentionalistic top-down metaphor: windows are opened by persons and under normal conditions they are opened deliberately. Naturalistic theorizing however, must operate on sub-personal levels of description and replace internal homunculi and their volitional acts with principles of self-organization. Besides, the time-window metaphor is one of those notorious spatial and visual metaphors which have bewitched the occidental philosophy of mind ever since it began. It suggests too easily a fixed view-point and distal objects. A naturalistic analysis of the formation of representational wholes however, requires a conceptually convincing theory of self-organization without a homunculus. The homunculus, the window, and the distal object easily creep into our scientific theories operating 'from the outside' as well. Since in verbal descriptions form and content again become separated, representational wholes turn into abstract objects: They are objects in high-dimensional representational spaces, the formal structure of which may be captured by mathematical abstractions such as activation vector space descriptions of a certain type or neural algorithms with certain time-constants. However, in the sub-symbolic process which I have called the 'opening of a time-window', the system itself can no longer distinguish between form and content: The 'opening of the window' and the 'generation of the object seen through the window' are identical, both descriptions are co-extensive. They refer to the very same activation pattern in the brain which is bound through self-organizing processes of synchronization: The neurally realized integration window is an object, a new state of the system. Like a flower, it opens by itself. For this reason, I shall abandon metaphorical talk about time-windows and will henceforth speak about 'functionally active representational wholes'.
I will now go through the assumptions Al to A4 and offer some short conceptual speculations for each of them. Naturally, in the context of the two explananda isolated at the outset of the paper, the assumptions Al and A4 deserve our attention the most.
Part
3