Go to Semiotic Review of Books Home Page
Go to SRB Highlights
Go to SRB Archives
This editorial appeared in Volume 5 (1) of The Semiotic Review of Books.
In the 1980s a crucial shift has occurred in cognitive science. As it became clear that the traditional symbolic paradigm could not keep its promises to develop artificial cognitive systems / models which are "really intelligent", an alternative class of models (re)appeared on the scientific stage: the connectionist or Parallel Distributed Processing (PDP) or the neurocomputational approach. This class of models chose an alternative basis for its simulations: instead of applying the traditional form of symbol manipulation, rule systems, frames, etc. the neural substratum and its dynamics are considered to be the foundation of cognitive processes. In this context the term cognition is not restricted to so called "higher processes", such as reasoning, planning, problem solving, etc.,but also includes perceptual and motor processes and the generation of all kinds of behaviour. In fact, the connectionists' claim is that all cognitive activities are based on neural dynamics regardless of their "cognitive complexity". This implies that a way has to be found of embedding, for instance, language or symbol manipulation into this neural framework since linguistic structures are no longer available as an ultimate representational substratum.
So, what is the representational substratum in (natural and artificial) neural networks; how do the latter process and represent knowledge; what are the differences from the traditional approach? In order to answer these questions we have to take a quick look at the basic structure of artificial neural networks: the basic idea is that all our cognitive abilities are grounded in neural processes and, thus, we have to simulate the dynamics of natural neural systems in order to generate "artificial intelligent behaviour". The connectionist approach takes this into account and tries to simulate neural processes on a rather abstract level - the basic functionality can be summarized as follows (for more details consult Rumelhart et al.,1986, Vol. I & II, or Hertz et al. (1991):the simulation consists of a network of units (representing the neurons). These units are connected to each other via weights (representing the synaptic connections).
So units can excite or inhibit each other via positive or negative weights. A unit x computes its activation by summing up the products of the particular weights and the activations of the units with which unit x is connected. This product, which is called the netto-input, is then transformed and scaled by a squashing/activation function which computes the actual activation/output of the particular unit. This activation/output acts as an input for the other units being connected to this unit. Each unit in the network executes this function in parallel. This leads to an effect called spreading activations, which means that patterns of activations are spreading through the network in the process of being integrated and transmitted in a parallel process.
We have to differentiate between three types of units: in the input units, the transduced environmental stimulus/state is represented as a pattern of neural activations which is passed on to the hidden units inside the network. They are called hidden units because they do not have direct access to the environment - they play a crucial role in the context of neural processing and neural knowledge representation, however. Finally, the activations- reach the output units, where they are transformed into an output signal. Depending on the structure of the environment the in/output signals are realized in different ways. In general there are two possibilities: (a) "linguistic environment" - the neural network is embedded in an abstract symbolic/linguistic environment in which the in/output is represented in symbols (i.e., symbols are coded as patterns of in/output activations and vice versa). (b) In the case of a physical/sensorimotor embedding, the neural network is coupled via sensors and effectors to its physical environment. It seems to be clear that the latter is the more realistic, and biologically and epistemologically more plausible approach. For those interested in the functioning of cognitive processes P.S. Churchland et al. 1992 give a good overview of current biologically plausible models. Nevertheless, most models and technical applications are embedded in a linguistic environment.
The interesting thing about (artificial) neural networks is their approach to knowledge representation and the implications for traditional views in cognitive science. The problem of knowledge representation is closely connected to the issues of learning/adaptation in neural networks: as the synaptic weights control the flow/spreading of neural activations they are heavily involved in the process of knowledge representation in neural systems. The spreading of the activations leads to externally observable (more or less successful) behaviour which makes the observer think that the observed neural system must have some kind of representation of the environment in order to generate such (more or less) appropriate behaviour. Learning/adaptation is realized as a process of slight changes in the configuration of the weights. This brings about a change in the dynamics of the spreading activations leading to a change in the externally observable behaviour which makes the observers assume that the "knowledge about the enviroment" must have changed in the neural system. As these changes have occurred in the synaptic weights they must be major players in the process of knowledge representation in neural systems. By controlling the flow of activations, the synaptic weights do not represent some kind of knowledge about" the environment but they rather embody knowledge (successfully) with the environment. The physical structure/architecture can be understood as some kind of embodied theory which has adapted phylo- and ontogenetically towards a certain configuration (Peschl, 1992). It is we, as observers, who interpret the behaviour and impute to the underlying neural mechanism some kind of knowledge about the environment which has a similar (depicting and linguistic) structure as our own impression of the environment. What we are forgetting is that this "depicting" and/or linguistic impression of the environment are themselves the result of constructive processes in our own neural system.
Thus, representation in the synaptic weights cannot be understood as a kind of mapping from the environment to the representational substratum (such as in the case of a symbol referring to some environmental state) - we have to look at it as a gradual and incremental process of adaptation which is continuously going on in the weights/architecture which are changed on a trial&error basis until some kind of equilibrium between the actually generated behaviour (for which the current configuration of weights is responsible) and the internally and/or externally defined criteria for "successful", adequate, or functionally fitting behaviour are established. The relation between "representations" and "repraesentandum" is not a relation of direct reference, but rather an indirect "generative" relation. In other words, representation is not defined over a mapping function, but via the criterion of providing a physical substratum (e.g.,a certain configuration of weights/architecture) which is capable of generating adequate behaviour accounting for the survival and reproduction of the particular organism. "Survival" is understood not only in the "physical" sense of, for instance, finding food, but also in the sense of cultural, social, linguistic, and even scientific survival.
These considerations about knowledge representation in neural systems are only one example of the dynamics and interdisciplinarity in the field of computational neuroscience. Computational methods and concepts enable an understanding of a problem which has implications for not only empirical neuroscience, but also for epistemology which, by definition, is concerned with the nature and development of knowledge. Especially in the field of cognitive neuroscience the computational approaches, simulation, etc. have become a very important method (besides "real" empirical studies). From a philosophy of science perspective this is an interesting development, as researchers do not investigate natural processes, the environmental structure/dynamics, etc. any more, but make abstract simulations based on mathematical models in their computers. P.S. Churchland et al. (1992) report interesting examples in which the computational approach or a simulation of neural system have predicted neural phenomena which afterwards have been empirically verified. In fact, in the extremely controversial field of learning, neural plasticity, memory, etc. computational neuroscience has provided numerous and powerful concepts and suggestions - they have the advantage that they abstract over the rich details of empirical neuroscience and (in most cases) focus on the system theoretic and other general level. This is also a possible link to epistemological questions: one cannot expect from a philosopher that he/she knows all the details tabled in empirical neuroscience; the knowledge, theories, and general concepts being provided by computational neuroscience are (a) accessible (b) better understandable for them and (c) have direct implications and relevance for epistemological questions.
It turns out that these approaches have also implications for a number of other disciplines concerned with cognitive processes (e.g., linguistics, psychology, or semiotics), if they are combined with concepts from artificial life (e.g.,Langton et al. 1992, 1993): cognitive processes and knowledge representation are not restricted to only a single cognitive system/brain, but their dynamics can be studied in the context of social interaction, of the artifactual/cultural environment, of the evolutionary, and of phylogenetic development, etc. In other words, cognition is not seen any more as an isolated phenomenon of a single cognitive system, but has to be understood in the full context of our cultural, linguistic, scientific, and social environment. Let us illustrate this by two examples.
(a) The combination of genetic algorithms and artificial neural networks: in biological as well as artificial neural systems, we face the problem that the knowledge which is acquired during the lifetime of a single organism is lost with its death. What is passed on to the next generation is only the very basic architecture and basic principles of the body/neural dynamics (e.g.,how to learn, how to develop the body structure, etc.) The individual body structure as well as the actual architecture of the neural system has to be developed individually by applying trial-&-error processes in the course of the individual ontogenetic dynamics. Belew (1990) showed in a simple, but very impressive simulation experiment that the individual success of the ontogenetic adaptation/learning process is indirectly passed on to the next generations (via a higher likelihood of reproduction). This is an interesting result as it confirms our intuitive idea that "smart" organisms can "guide evolutionary" processes to a certain extent. (b) Hutchins et al. (1992) went one step further (and this could be of interest for semiotics): they introduced an artifactual structure (i.e., a simple form of symbols) which can represent the knowledge being acquired by an individual organism during its lifetime (of course, on a very simple level). i.e., the particular organism extemalizes its ontogenetically developed knowledge in the form of simple symbolic artifacts - they describe the "experiences" which this organism has made during its lifetime. These artifacts are accessible to all other organisms of the population. This implies that these organisms do not have to learn from direct environmental experience, but can learn from these artifacts The price they have to pay is, however, that they first have to learn some kind of language or symbol system (Peschl 1993 gives an epistemological account of how these processes can be realized in neural systems). Thus, Hutchins exactly simulates these processes on which our so called cultural life, our education, our science, etc. are based. It can be shown that by learning from artifacts, the development of "new" knowledge is speeded up as it is not necessary any more for each organism to physically go through all the experiences its predecessors have made.
What has been sketched by these two examples is that cognitive science and artificial neural networks finally begin to develop a community of cognitive systems, and that the introduction of social, cultural and phylogenetic interaction opens up a new dimension and might offer alternative explanations of phenomena which traditionally have been explained in rather speculative terms. The recent development of combining artificial neural networks with concepts from artificial life aims not only at disciplines and theories which are concerned with neuroscience, but also at disciplines studying language, general symbol systems, anthropological issues, social interaction, etc. For them, the challenge is to accept these alternative, "neurally grounded" explanations and their implications, and to integrate them into their traditional theoretical frameworks.
Belew R.K. (1990), "Evolution, Learning and Culture: Computational Metaphors for Adaptive Algorithms." Complex Systems 4,pp 11-49.
Churchland P.S. & Sejnowski T.J.(1992), The Computational Brain. The MIT Press, Cambridge, MA,
Hertz J. Krogh A. & Palmer R.G. (1991), Introduction to the Theory of Neural Computation, Addison Wesley, Redwood, CA,
Hutchins E. & Hazelhurst B. (1992), "Learning in the Cultural Process." In C. Langton et al. (eds), Artifical Life II, Addison-Wesley,CA, pp 689-706.
Langton C.G.,Taylor C., Farmer J.D. & Rasmussen S. (eds.) (1992, 1993) Artifical Life lI;Artificial Life III, Redwood City, CA, Addison Wesley.
Peschl M.F. (1992), "Construction, Representation, and the Embodiment of Knowledge, Meaning, and Symbols in Neural Structures: Towards an Alternative Understanding of Knowledge Representation and Philosophy of Science," Connection Science, Special Issue: Philosophical Issues in connectionist modelling, Vol 4, Nos 3 & 4, pp.327-338,1992.
Rumethart D.E. & McClelland J.L. (1986), Parallel Distributed Processing, Explorations in the Microstructure of Cognition, Volume I & Il, MIT Press, Cambridge, MA.
Dr. Markus Peschl teaches in the Department of Philosophy of Science, Unit of Epistemology and Cognitive Science, at the University of Vienna (Austria). He is currently a Visiting Scholar at the University of California, San Diego. He is the author of several books (in German) and many articles on cognitive modelling, subsymbolic neural representation systems, connectionism and evolutionary epistemology.