| Sažetak | Složeni inteligentni sustavi za uspješno rješavanje problema iz stvarnog svijeta zahtijevaju veliku količinu znanja i učinkovite postupke zaključivanja. Znanje o problemima iz stvarnog svijeta je neizrazito, nejasno i dvosmisleno. Za rješenje navedenih problema predlaže se biološki, neurološki i psihološki inspirirana arhitektura hijerarhijske heterogene baze znanja. Arhitektura omogućava izgradnju složenih inteligentnih sustava prikladnih za razna područja primjene. Baza znanja ima tri razine: asocijativnu, semantičku i razinu s pravilima. Za bazu znanja razvijeni su ulančani paralelni neizraziti postupci zaključivanja koji koristite znanje pohranjeno na trima razinama koje koriste različite sheme za predstavljanje znanja. Za učinkovito paralelno raspodijeljeno pohranjivanje i pretraživanje velikog broja pojmova u asocijativnoj razini, na temelju sličnosti opisane neizrazitom lingvističkom varijablom, razvijena je prikladna arhitektura asocijativne memorije. |
| Sažetak (engleski) | Complex intelligent systems require a huge amount of knowledge, as well as some efficacious mechanisms for manipulating this knowledge, in order to perceive, organize and summarize stimuli obtained from the real world. They also require a large amount of information from a problem domain and efficient reasoning procedures to support a useful degree of problem-solving ability which is limited by three major obstacles. The first obstacle is complexity of world knowledge which humans acquire through experience, communication and education. The second obstacle centres on the concept of relevance which is very complex. The third obstacle is deduction from perception-based information which is fuzzy, uncertain and ambiguous. Architecture of a hierarchical heterogeneous knowledge base suitable for solving these problems is proposed in this dissertation. The architecture enables development of complex intelligent systems suitable for various problem domains. Motivation and aim of the research is to develop a knowledge base model capable for efficient reasoning with human knowledge about real or abstract concepts obtained from the real world that are fuzzy, uncertain, vague and ambiguous. The model is inspired by biological, neurological and psychological models obtained by analyzing how human and animal brains abstract, process and store knowledge from the interaction with the environment. A brain has hierarchical heterogeneous structure. It consists of distinct layers of neuron cells which are hierarchically stacked in brain’s tissues and interconnected with forward and backward connections. These layers are formed from characteristic cell types of neurons. Biological and neurological models of the human and animal brains show clear distinction between processing of familiar well known stimuli and novel stimuli during an interaction with the environment. The integration within and among specialized areas is mediated by forward, backward and lateral connections and is very complex. The important anatomical and functional distinctions between forward and backward connections are: forward connections are less divergent, and transmit known stimuli directly to higher levels, and backward connections are more divergent and they are used when processing unknown stimuli. Familiar environmental stimuli travel fast by means of the forward connections from layers of neuron cells in sensors to top layers of neuron cells in a brain’s cognitive system. Novel stimuli at one layer of neurons create an “unknown” neuron activity patterns that are unable to activate neurons at higher layers with forward connections (i.e., input stimuli are not “recognized”). In this case backward connections are used to activate large areas of a lower layer in the hierarchy of the layers in the brain. Consequent neural activity patterns at the lower layer result with multiple activations of forward connections corresponding to some similar known stimuli that are connected in the experience and have similar neuron activity patterns with the stimulus that is unknown at the upper layer. This process of activation is based on the maximum overlap of subsets of neurons that encode known and unknown stimuli. A minimum overlap which is required to recognize an incoming stimulus as a known, already learned, concept is not exactly known but it has to be significantly larger than an overlap between two random stimuli. This process enables the brain to process novel stimuli based on similarity estimations to already known concepts obtained from lower levels. In short, novel unknown stimuli at one layer are substituted with most similar stimuli obtained from a lower layer. Similarity between stimuli is automatically established based on an overlap between corresponding neural activity patterns. If some of the obtained similar stimuli are known at the upper layer than cognitive reasoning is able to continue, if not, than the procedure is recursively repeated with similar concepts retrieved in the previous step. This forward and backward connections multilayer interplay is recursive with many cycles and can last very long time for novel situations and difficult cognitive problems (i.e., seconds, days, or longer). Familiar situations and well learned stimuli can be successfully interpreted just after few milliseconds. This can explain why familiar faces are recognized in milliseconds due to dominant use of forward connections but recognition of faces of distant friends or relatives can take minutes, or even longer, due to use of the backward connections recursively in many cycles. This dissertation describes the knowledge base model inspired by the above described organization of a brain, as well as the underlining reasoning processes. Stimuli learned through the interaction with the entities from the environment are modelled as concepts. Entity is something that exists in the real or an abstract world separately from other things and has its own identity. The entities are perceived with percepts. Percepts in general are basic component in the formation of a concept. Entities are represented as concepts. A concept is an idea or a principle that is connected with an entity. A concept is described with one unique term (a word or linguistic expression) and with one or more percepts (pictures, sounds, or any other perception-based representations). Each percept can be characterised with its modality (visual, auditory, infrared, radar, ultrasound, ...) and with its specific type (shape, colour, video, face image, iris image, pitch, tone, voice sequence, song, ... ). The proposed architecture of the hierarchical heterogeneous knowledge base has three levels: an associative level, a semantic level and a rule-based level. Architecture of an associative memory suitable for efficient parallel distributed storing and searching a large number of concepts at the associative level, based on similarities expressed with values of a fuzzy linguistic variable, is developed. The associative level is composed of several modality modules, each of which is composed of several percept modules. The levels use different knowledge representation schemes which support reasoning procedures similar to cognitive processes in humans and animals. The levels in the knowledge base model are hierarchically organized and connected with forward and backward connections. Initialization of reasoning procedures performed at the levels in the model is equivalent to neuron activations mediated by forward connections in a brain. Neuron activations mediated by backward connections are modelled with an activation of reasoning procedures at a lower level. Neuron activations mediated by lateral connections at neuron layers are equivalent to reasoning processes performed at the three levels in the model. Cycles of forward and backward connections activations in a brain are modelled as chaining of reasoning procedures. The model is able to replicate cognitive processes. |
