Ассоциативная память — среда для формирования пространства знаний. От биологии к приложениям.
The monograph deals with the analysis of the structure and functionality of the human brain, the results of which are presented on the example of processing quasi-linguistic information of various types using operating software models. It also describes the implementation of an element of corticomorphic associative memory developed on the basis of these results, in the form of a microelectronic model – associative memory VLSI with fuzzy search.
Three basic structures of the brain: the cortex of the large hemispheres, the hippocampus and the thalamus are responsible for processing information in the brain obtained from the sensory organs of a human. Columns of the cortex form dictionaries of images for events of varying complexity and various modalities. The cortex together with the hippocampus forms representations of whole situations in the lamellae of the hippocampus on the basis of event images formed and stored in the cortical columns. The hippocampus, along with the cortex, reorders the presentations in the cortical columns involved in representing the situations in the hippocampus lamellae, while simultaneously ranking the weight characteristics of the event images in the cortex. In other words, in the cortex and the hippocampus, a structural representation of the external and internal world of a person is formed, which has the property of associative access to information. The thalamus, as a control structure, by controlling all information flows layer by layer at the inlet, outlet and inside the cortex, allows control of the processes in the cortex strengthening some of them and weakening others.
The cerebral cortex column, as an association of pyramidal neurons of the third layer, performs structural processing of specific information modeling a fragment of a multidimensional signal space. The model is implemented by an artificial neural network based on neuron-like elements with time summation of signals. Input sensory information is a matrix of information flows, where each information flow can be considered as a quasi-text consisting of repeating elements of various complexity and with various frequency of occurrence. The brain is a parallel computer that implements layerwise parallel processing of information coming from the sensory organs through the subcortical nuclei, in the cortex, and again through the subcortical nuclei to the effectors.
The column of the cortex modeling a fragment of a multidimensional space performs the transformation of the incoming information flow into a trajectory in a multidimensional space. In a multidimensional space, in the process of structural processing of input quasi-texts of various modalities, dictionaries of images are formed for events of the internal and external world of various complexity, which make up a multi-level hierarchical structure. Information relations between the levels allow filtering out the words of the current level from the input quasi-text, sending further the new information characterizing the relations of these words in the input information, which at the next level can also be used to form a dictionary of the next level. As a result of this processing, dictionaries of event images of various levels are formed, from the simplest to the most complex elements in this modality. If we consider the text of a natural language as an example, then these dictionaries form images of events of morphological, lexical, syntactic, and, finally, semantic levels. The semantic level in this case means the level of representation of the semantics of a separate sentence, where the lexical component of the semantics of the sentence is represented by the pairwise compatibility of concepts (words or word combinations).
The pairwise compatibility of concepts allows formation of a semantic network of the whole text. In the brain the so-called cognitive semantic networks are formed, networks that are built on a physical substrate - a natural neural network, the brain. These networks are built on the principle of accounting for the joint occurrence of individual images of events in situations, that is, the semantic relations between the concepts in these networks – of only one type, associative ones. Therefore, cognitive semantic networks formed are homogeneous (that is, associative). The primary semantic network is characterized by frequency weight characteristics.
The information constantly added during the learning process requires its correlation with the previously processed information already stored, for which constant corrective weighting of the parameters is necessary as the vertices of the network and their connections. The role of the novelty filter is played by the hippocampus. In its lamellae, information is formed and stored on relations of the images of events stored in the cortical columns, as they appear to be connected in various situations. The re-ranking of network concepts is performed by an iterative procedure similar to the Hopfield network algorithm. The structure of representations of situations in the hippocampus lamellae resembles the structure of the semantic graph of a separate sentence of the text, the extended predicate structure of the sentence.
In the cerebral cortex, hierarchies of modal dictionaries are formed; at higher levels of representation a hierarchy of multimodal dictionaries is formed. The collection of all the words from all dictionaries of all levels of all modalities, as well as all the words from all dictionaries of all levels of the multimodal representation, represent a static model of the human world in which information about the world is stored in terms of images of events of various complexity, and which interact with each other according to the principle of associative reference to information: similar information is stored in close dictionaries. That is why the whole model of the world is a semantic network in which three subnets are prominently distinguished: linguistic; multimodal multilevel schematic, and multimodal multi-level image bearing (individual) ones. The formation of dictionaries of the multimodal multi-level representation is performed bottom-up with gradual filling of levels of more and more complex representations. The world model is clearly structured into three main components: two in the dominant hemisphere, and one in the subdominant one. In the dominant hemisphere, the linguistic part of the world model is formed through the language under the influence of society, as well as a multi-level multimodal schematic part of the world model. While in the subdominant hemisphere a multimodal two-level individual part of the world model is formed, which is less dependent on the society and mostly takes into account the individual history of the development of a particular person. Due to the level-by-level links of the linguistic and multimodal parts of the world model, and also due to the synchronism of their formation, the language model of the world is isomorphic to the multimodal multi-level schematic part of the world model in the dominant hemisphere.
If we consider a particular quasi-text projecting it onto the semantic network of the world model, then we can see that it is projected into a chain of words of dictionaries, which corresponds to a specific description of a certain subject domain. This chain can be transformed into a chain of extended predicate structures of the sentence corresponding to the fragment of a quasi-text representation of a particular modality described by this sentence (for example, a video sequence). A lot of such chains can be structured, on the one hand, into two classes: descriptions and algorithms, and on the other hand (already within the framework of a particular subject domain) into classes of structurally close chains that can be replaced by one representative of the class. Such (on the basis of many classes of chains) a representation is naturally called pragmatic.
On the example of the formation of the sensory information flow of the visual modality in the precise channel of the human visual analyzer, it can be shown that all sensory flows are quasi-texts that resemble natural language texts in structure. Therefore quasi-texts, like natural language texts, can be processed as natural language texts using the structural approach. Since the linguistic methods of text processing based on formal signs of the text form allow processing information up to the syntactic-semantic level of a single sentence, the formation of a semantic representation of the text is possible only through statistical methods [Kharlamov, 2006]. Algorithms of the structural statistical analysis of the semantics of the whole text are presented on the example of the work of the program neural network technology TextAnalyst. The TextAnalyst technology is implemented on the basis of the formation of dictionaries of event images in the columns of the cortex, which are a homogeneous semantic network, as well as the re-ranking of the weight characteristics of the vertices – concepts of this network, as it is done in the joint work of cortical columns and the hippocampus lamellae. The main functions implemented within the TextAnalyst technology are the formation of a semantic network, re-ranking of weight characteristics of concepts, formation of a summary, comparison of texts in meaning (by structure) and classification of texts. The TextAnalyst technology is the base for creating e-books that allow non-linear text reading.
Since the corticomorphic associative memory is an effective tool for analyzing any quasi-text information, its microintegral implementation can be useful for creating information processing hardware. Modern computer technology is at the limit of applicability of well-known Moore’s law, since the design standards of microelectronic technology have come close to the size of the atom, when absolutely different laws of physics begin to work. Objectively, the time has come for the transition from the traditional sequential von Neumann paradigm of building computing devices to other approaches. One of such new ways of development of computer technology can be the use of the corticomorphic associative memory for the construction of massively parallel computing structures that use the “Correlation-Extremation-Uncertainty” computational basis, rather than the “AND-OR-NOT” traditional basis.