J. Kumičák
Department of Thermodynamics, Technical University,
Letná 9, 041 87 Ko�ice, Slovakia
Abstract
The process of learning occurring during the life of an individual can be viewed as an analogy to phylogenesis of species. This approach enables one to apply evolutionary ideas and principles to learning in the most direct way. The role of selective agent is played here by consciousness and it turns out that in this phenomenon the nature has discovered a mechanism to predict the future. A model of a kind of prediction-oriented learning is discussed shortly with emphasis on its simplicity. It seems namely that the desired simplicity can be achieved presumably only when simulating the phylogenesis of the corresponding algorithm.
The approach to irreversibility elaborated by the Brussels school under the guidance of Professor I. Prigogine [1] is deeply connected to what they sometimes call the arrow of time. This term denotes the invariably observed fact that evolution of macroscopic physical systems prefers one direction of time, called future. We encounter this preferred orientation not only in evolution of macroscopic systems, to which statistical physics applies, but also in such fundamental processes as is the scattering of quantum particles [2]. No wonder that we observe the same orientation to future in living systems as well since they are equally subject to physical laws. What is, however, really exciting is that this phenomenon is encountered also in the mysterious phenomenon of consciousness where it actually culminates.
This makes the problem of consciousness — the former mind-body problem of philosophers — the topic of interest of modern physics [3]. However, it seems that the relation of the problem to the arrow of time remains unnoticed in all such investigations. The aim of the contribution is to show that such a connection is at least plausible.
All nature seems to be governed by the past: the present states determine the future ones and this relation is usually called causality. Conscious subjects, however, make decisions based also on their wishes. They represent what is called intentional systems — systems having beliefs, desires and expectations [4]. Their behaviour is frequently determined by the anticipated result — by the imagined future. It looks as if the causal link were reversed here: first comes the desire, which is then followed by the behaviour leading to the expected outcome. In other words, the majority of systems are "pushed" by the present state into unknown future, whereas conscious subjects are "pulled" by the expected and imagined future. This suggests that in conscious subjects the future plays the constructive role.
If we conceive of the future as of the cognized possibility subjected to consciousness, we may say that it did not exist sooner than consciousness had appeared. It entered the scene only when the phylogenesis brought into being creatures that were able to imagine possible future states and tried either to avoid them (as in case of danger) or approach them (for example to reach food). Consciousness is a very specific way of mapping the outer physical world into inner mental states of a subject. It is an operation which has an "inverse" able to map mental states into possible physical events. This ability culminates in man's creative thinking when the mind invents the mental image of a thing which has never existed before. The thing is then created and can be perceived by conscious subjects. This observation supports the view that subjectively perceived future is an objective state of subject's nervous system and as such it exists objectively.
Future emerged in the course of evolution of nervous system — and especially of the brain with its neural structure — when the ability to learn had appeared. Artificial learning can be, in principle, explained even without introduction of consciousness, but the learning observed at least in higher animals relies on consciousness heavily. That is why we will not consider the possibility of learning in systems lacking consciousness. Learning can be then described in the following way. Within individual life spontaneous mutations in behaviour occur which are rewarded by pleasant feeling or punished e. g. by pain (a process known as operant conditioning). Both rewarded and punished behavioral patterns are remembered but only the rewarded ones are activated later upon the presentation of adequate stimuli. Whereas in phylogenesis survival was the "reward", in ontogenetic evolution of learned behaviour the role of reward is played by satisfaction brought about by success. The emergence of ability to learn can thus be formulated in such a way that the mechanism of evolution has moved from phylogenesis into ontogenesis: to learn means to apply selection to behavioral patterns. This accelerates the process of evolution essentially: whereas in phylogenesis there is time delay of a generation between mutation and its reward, in learning this is shortened to seconds.
In phylogenesis the supervisor was blind and deaf — it was the unconscious nature. In learning the supervisor is consciousness itself. Hardwired reactions were stereotypes of the stimulus-reaction type. Learned behaviour can encompass even reactions to events which have not happened yet. In consciousness the nature has discovered a mechanism to predict the future.
Of paramount importance is the question of possibility to model conscious behaviour by an artificial system, such as computer. The first formulation of the question [5] opened the way to attempts aiming at its solution. One of the alternatives resides in the theory of neural networks. There are many rather efficient models of learning (usually tailored to fit specialized tasks). However, from the above remarks it follows clearly that a sensible approach to the solution of the problem of consciousness based on neural network models of cognitive systems can be expected only if the networks incorporate image of future, only if the problem-solving starts from the "mental" picture of the goal. We describe briefly the basic structure of a model which is able, in principle, to simulate at least basic treats of behaviour which could be denoted as prediction-oriented learning.
The model consists essentially of three interconnected components accomplishing the following tasks: a) generation of intentions, b) problem-solving which generates the behaviour leading to fulfilment of the intention starting from "mental image" of the goal and c) evaluation of behavioral "trajectories". Obviously, in living beings, the component c) is the result of phylogenesis (systems, evaluating self-destruction as success, would be quickly eliminated) and it can be implanted to the model from outside. The generation of intentions is also partly the result of evolution but the image of an intention is dependent on individual experiences so that it is the result of learning and remembering and has to be simulated by a kind of associative memory. The real challenge represents the problem-solving based on mental picture of reality. The mapping of reality by consciousness must ensure a kind of isomorphism between manipulations with the picture and manipulations with real objects. The model must allow for deformation, combination, exchange, etc. of components of the imagined behaviour. Each behavioral trajectory is then evaluated and the best rated one chosen for realization.
One thus sees that the "mental" problem-solving saves energy which would be otherwise spent on real manipulations and minimizes eventual fatal outcomes. Generation of variants of behavioral trajectories and their evaluation must be based on learning — maybe primitive — of cause-and-effect relationships. The predictive force of the system thus originates in its ability to find a variety of realistic behavioral patterns and choosing the "best" of them. There is then no real time reversal of cause and effect — the system simply traces as many trajectories as possible and makes a choice based on its individual experience. This is the sense in which consciousness can be said to predict the future.
The nature has discovered the most essential constituents of consciousness evidently at very early stages of evolution of higher animals (as evidenced by the spread of the phenomenon among all vertebrates) so that basic structures must be relatively simple. The most fundamental problem to achieve in designing a realistic model is therefore its simplicity. It seems that one can approach it only simulating the "phylogenesis" of the model using some of genetic algorithms [6]. Though the task is not easy, we are left with no other choice than to search for the essence of prediction of future since it is there where the arrow of time is pointing.