V.G. Yakhno

We propose the recognition-system architecture which permits formalized description of different cognitive functions of the live systems. For description of specific features of the cognitive functions, the following units of subsystems were singled out: a) recording of sensor (visual, auditory, chemical, etc.) signals, calculation of their features and filtering masks; b) variation in the set of algorithms for processing and visualization of separate objects in these signals; c) variation in the set of algorithms for processing, calculation of comparison measures, and decision-taking accuracy; d) variation in the capabilities of algorithms making it possible to start imitation of input sensor signals by using a code description of system states in the past and assumed state in the future. The system should provide for joint operation of the coding subsystem and an additional subsystem of "inner imitation" (reconstruction, virtual presentation) of the expected input sensor signal. Such an architecture of signal processing ensures cyclic (recursive) processes that ensure the choice of the most appropriate filtering mask on the input image or a more adequate model for analyzed object as well as calculation of additional features for obtaining "integral" estimates which can be put into correspondence with the signal "conscious perception" process. The elementary model of "conscious perception" of the input-signal image in such a model system is related to the dynamic coding - reconstruction process (generation of interpreted signal) and search for optimal algorithms in the system. Unconscious processing of the input signal corresponds to the cases of absence of cycles related to generation of interpreted signals in the system. It is possible to consider the hypotheses of "intuition" mechanisms (solutions taken under conditions of absence of previous experience in the current problem), which in some papers are confused with the unconscious mode of behavior (unconscious decision taking for well trained objects). The process of focusing on expected image in such an "elementary adaptive recognition system" corresponds to the choice and "loading" of a package of models (algorithms) which are necessary for its work with expected image. Significance of the input informative signal can be determined from the values of the differences (errors) calculated from a comparison of sets of code descriptions for the image preliminarily expected by the system and the actually calculated codes from the input image. The same values of the differences serve as a stimulus for the motivation formation and decision taking with respect to the image chosen by the system. The efficiency of the presented definitions was demonstrated with the use of neuron-like detectors, which we developed, for real-time detection of the object chosen beforehand against an arbitrary complex background. The detector formed in such a way uses sets of features from different spatial regions, which are calculated with the help of operations similar to transformations in visual receptive fields of live systems. It is shown that the psychological and neurological human states can be described in a consistent way provided that the parameters determining the basic states of the corresponding model system (simulator) are recorded.