Search

The class of self-organizing systems represented by networks which learn to recognize patterns is reviewed from an historical standpoint, and some of the behavioral similarities between such nets and biological nervous systems are discussed. Examples and results of several experimental models for alphanumeric character recognition are presented. The network synthesis problem is then recast in terms of redundant information removal, multivariable curve-fitting and expansion in orthonormal functions. Recognition network structures and the learning process are described from these points of view. The potential component and behavioral advantages to be gained from sequential feedback networks are discussed briefly.

A general discussion of ideas and definitions relating to self-organizing systems and their synthesis is given, together with remarks concerning their simulation by digital computer. Synthesis and simulation of an actual system is then described. This system, initially randomly organized within wide limits, organizes itself to perform a simple prescribed task.

The development of a system’s biology, as a common construct for both physiologist and engineer, requires both a theory of structures (form) and a theory of dynamics (function). A dynamic organizing principle—“homeokinesis”—for the living system was proposed earlier. Based on thermodynamic reasoning, homeokinesis attempts to capture the physical essence of homeostasis. Now, a primitive foundation is proposed from which a large family of design characteristics might emerge, by self-organization, in complex biological organisms. This foundation is directed at the emergence of major form parameters of the entire class of mammalia, from 3 gm adult shrews to 100,000 kg whales.

A nerve net model for the visual cortex of higher vertebrates is presented. A simple learning procedure is shown to be sufficient for the organization of some essential functional properties of single units. The rather special assumptions usually made in the literature regarding preorganization of the visual cortex are thereby avoided. The model consists of 338 neurones forming a sheet analogous to the cortex. The neurones are connected randomly to a “retina” of 19 cells. Nine different stimuli in the form of light bars were applied. The afferent connections were modified according to a mechanism of synaptic training. After twenty presentations of all the stimuli individual cortical neurones became sensitive to only one orientation. Neurones with the same or similar orientation sensitivity tended to appear in clusters, which are analogous to cortical columns. The system was shown to be insensitive to a background of disturbing input excitations during learning. After learning it was able to repair small defects introduced into the wiring and was relatively insensitive to stimuli not used during training.

Virtually all self-organizing systems proposed as models for learning require a reinforcement control device. The biological analogue of this device has never been clear. Slow changes in the cortical potential may be correlates of a biological analogue of reinforcement control devices. Frequent pairings of novel stimuli with other stimuli eliciting specific responses produce a slow shift and oscillation in the cortical potential. This D. C. shift and oscillation occurs first in the sensory, and then in the motor cortex involved; it may be associated with the decreases or increases which occur in the threshold of cells in these cortical areas.

Several alternative systems of "adjustable logic" are available for use in a self-organizing system having discrete channels as inputs and outputs. For example, a device might consist of a network of McCulloch-Pitts neurons with variable thresholds, or use "summation logic" (refs. 1, 2) with variable weights attached to the input signals. A further possibility is described by Kochen (ref, 3). The system of "adjustable logic" used in the present study depends on the evaluation of a polynomial function of the input signals, with a facility for adjustment of the coefficients of the polynomial terms. The output is a single binary signal of value 0 or 1 depending on whether the polynomial sum is negative or positive. Means are provided for introducing new terms if certain criteria show them to be advantageous, and of eliminating terms which are redundant. It is trivial to show that any logical function whatever can be represented by a polynomial in this way. This system of logic was originally selected because polynomial functions can also be employed in self-organizing systems for continuous variables. (refs. 4, 5, 6, 7). The system with discrete signals was intended to illustrate certain strategies which are applicable to the continuous case. Important differences, however, do exist between the discrete and continuous cases, and further work on continuous systems is in progress. Nevertheless, the work on discrete systems has produced a number of interesting conclusions. The facility whereby extra terms can be added to the polynomial function is similar to the capacity for amoebalike growth shown by the "Janet" program due to Foulkes (ref. 8).

THE most rational way of constructing a theory of discrete self-organizing systems seems to be on the basis of the structural theory of automata (see [1] or [2]). A self-organizing system is represented in the form of a logical network having a variable structure and sometimes with a number of elements varying with time. Naturally by remaining on the level of the abstract theory of automata we lose the large number of possible definitions. The present paper attempts to show that, in spite of the qualifications mentioned and within the limits of the abstract theory of automata, it is possible to construct a satisfactory theory of self-organizing systems.

This paper considers an aspect of the problem of representing the process of self-organization by a feedback control process. In particular, the problem is considered where the desired objective of the self-organizing system can be attained only by first setting up certain intermediate sub-goals. These sub-goals are continually adjusted by giving them varying weighting factors. Here, the problem of determining these weighting factors is considered. The approach is by the use of contraction mapping as a method of representing a feedback control process.

The paper makes an attempt to resolve two conceptual mingling: (a) the mingling of the two interpretations of the concept of orderedness applied in statistical thermodynamics and in symmetrology, and (b) the mingling of two interpretations of evolution applied in global and local processes. In conclusion, it formulates a new interpretation on the relation of the emergence of new material qualities in selforganizing processes on the one hand, and the evolution of the universe, on the other. The process of evolution is a sequence of emergence of new material qualities by self-organization processes, which happen in negligible small segments of the universe. Although thermodynamics looks at the universe as a closed (isolated) system, this holds for its outside boundaries only, while the universe has many subsystems inside, which are not isolated (closed), since they are in a permanent exchange of matter, energy, etc. with their environment (with the rest of the universe) through their open boundaries. Any ";;emergence";; takes place, i.e., all new qualities come into being just in these small open segments of the universe. The conditions to apply the second law of thermodynamics are not present here. Therefore, global evolution of the universe is the consequence of local symmetry decreases, local decreases of orderedness, and possible local decreases of entropy.

Summary 1. The Position of the Problem. 2. The Perception: A Machine for Recognition Without Categorization. 3. A General Categorization Method. 4. An Example of Categorization of Unidimensional Patterns.

Self-organizing systems are investigated and the basic foundations of a new approach arc proposed. Mainly based upon the notion of information exchange, this tentative defines the physical entropy of a system and some other characteristic functions. It is shown that a system which is isolated from its environment cannot increase its organization.

A self-organizing procedure to achieve a performance adaptive controller with asymptotically optimal properties is proposed for systems with completely or partially unknown dynamics. A physically realizable controller that operates in an unknown stochastic environment is obtained. The accrued cost during a global random search for the minimum converges to the minimum value corresponding to the specific optimal controller. The global search strategy includes a subgoal defined on a nondecreasing time subinterval and an algorithm of adaptive random type. The stochastic fuel regulator problem with random switching delay is used as an application of the method and simulation results demonstrate its effectiveness.

Various information-processing capabilities of self-organizing nets of threshold elements are studied. A self-organizing net, learning from patterns or pattern sequences given from outside as stimuli, "remembers" some of them as stable equilibrium states or state-transition sequences of the net. A condition where many patterns and pattern sequences are remembered in a net at the same time is shown. The stability degree of their remembrance and recalling under noise disturbances is investigated theoretically. For this purpose, the stability of state transition in an autonomous logical net of threshold elements is studied by the use of characteristics of threshold elements.

Cloud cover photographs transmitted from meteorological satellites must be processed and interpreted before weather maps can be issued. Most of the routine processing can be handled by present day digital computer techniques; however, the recognition and interpretation of cloud patterns such as vortices indicating hurricanes, must still be performed by humans due to the lack of suitable recognition mechanisms. This paper investigates the feasibility of using a perceptron-type computer for the recognition of vortex patterns. A formula is derived which enables the prediction of machine performance as a function of problem complexity and perceptron size (number of logic units). It is shown that the problem complexity can be estimated through optical correlation measurements on cloud cover negatives. These measurements are described and a computer routine is developed which mechanizes the prediction equations and examines the experimental data gained from 10,000 measurements. The results of the computer program are presented and their meaning is discussed.

Most modern theories of behavior involve a hierarchical structure whereby low level behavioral units are controlled or manipulated by higher centers so as to produce characteristic patterns of movement. In the simplest life forms, low level behavioral units may consist of simple reflex arcs with very little higher level control. In intermediate forms, the low level behavioral units may be relatively complex in themselves and subject to sophisticated control from higher centers. In the most advanced nervous systems, higher centers may themselves be arrayed in a hierarchical structure, with each level monitoring activity and exerting control over the levels beneath it.

Publisher Summary Brick and Zames showed how Wiener's expansion procedure could be used to synthesize decision functions for stochastic inputs. In this chapter, it is accomplished by adopting Ashby's approach to intelligence. Thus creativity, planning, induction, and hypothesis generation are features, which will not specifically be a concern for the present model of the automaton. However, generalization of existing hypotheses arid selection, recognition and learning with respect to these are major factors in the present chapter. The model of the recognition system derived here is, in fact, a form of conditional probability computer, which operates on stochastic processes. Of major interest is the fact that this procedure appears to avoid the tendency toward exponential growth normally considered to be characteristic of conditional probability computers. The narrowing-down of the artificial intelligence problem to the recognition, learning, and prediction phases, however one proceeds to justify it, is not without precedence.

AN ADAPTIVE MEMORY may be conveniently viewed as a space in which data are stored and which is subjected to structural deformation as a function of the ways in which the data are used. The intent of the deformation is to produce a space whose metric places in close proximity information which is highly correlated historically. The average access time per piece of data in a data request may be fairly low even though access time to the first piece located is large. After enough data requests have been serviced so that the statistics of the environment seem to be stable—and thus the space is holding still—a second phase of adaptation may occur. In this phase an attempt is made to generate two functions which map from the formal representation of the data request and the data, respectively, into the space. These are referred to as the retrieval function and the storage function, respectively.

Based upon the paradigm of self-organization, the present paper investigates the concept of cognitive memory from a non-representationalist perspective. Basic principles of self-organization are presented to provide elements for an ecological approach to memory. This is contrasted with the traditional Cartesian notion of memory which is intrinsically connected to mental representation.

A model of a simple self-organizing system is investigated. Mechanisms that are important to a molecular-biological understanding of evolution are studied with the aid of the model.

A self-organizing system, which may be biological or man-made, adjusts itself in response to inputs from the surroundings. The input information is processed and transformed, so as to guide the system in accordance with a desired final state.