Human Brain Inspired Cognitive Computational Model Based Robot Visual Learning Method

Human-like intelligence has been the objective in the field of artificial intelligence androbotics for long. Currently, mature and successful applications of robots are mainlytask-specific in a certain environment. With the growing demand of intelligence, robots arerequired to be highly autonomous and closer to human-like intelligence so as to serve fortask-nonspecific, in a highly complex environment. However, there are disadvantages andlimitations in traditional artificial intelligence method when designing robot intelligence system,such as the specification of tasks, off-line learning, poor expansibility of intelligence and badadaptivity in uncertain environment or tasks. Inspired by cognitive science, neurobiology,psychology and other interdisciplinary subjects, a new research area named cognitive robotics isproposed by researchers to solve the problem caused by the disadvantages and limitationsreferred to above. In this thesis, research work and contributions are done based on this risingarea of cognitive robotics. Our own method is proposed to try to make a step forward in thisarea.By learning from the theories of human brain cognitive mechanism in cognitive science andneurobiology, a cognitive computational model is built based on hippocampus-prefrontalmemory system. Visual perception autonomous learning algorithms are proposed based on thismodel to simulate the cognitive mechanism of information processing and learning procedure inhuman brain. With this method, the robot is able to autonomously explore its surroundings, learnvisual knowledge online, adjust learning strategies through the continuous interaction with theenvironment, and accumulate experience to develop intelligence continuously.The main contribution of this thesis is as follows:(1) Based on a thorough understanding of the working mechanism of human brain incognitive science and neurobiology, a cognitive computational model based onhippocampus-prefrontal memory system is proposed. The model is composed of sensorymapping, working memory and growing long-term memory, which simulate the memory structure and function of human brain and realize complex information processing mechanism.(2)Aiming at the problem of mapping high dimensional visual perception to lowdimensional representation, an online PCA with adaptive subspace algorithm is proposed to maponline the visual perception into internal representation for further processing by memory system.In this algorithm, the subspace updating strategy is adjusted adaptively according to the degreeof difference between new samples and learned samples. This algorithm is able to not only mapvisual perception to internal representation online but also recognize visual objects and scenes.Experimental result shows that the algorithm is able to cope with unknown sample problems andrealize online perception, accumulation and the update of visual inputs. As a result, theperception and recognition ability of system is gradually enhanced.(3)The mathematical description of hippocampus-prefrontal neural loop based workingmemory is proposed. The prefrontal working mechanism is described by dopamine controlledQ-learning. The hippocampus is described by visual novelty internal motivation, which drivesand controls the prefrontal Q-learning, to simulate the function of working memory in brain. Avisual novelty driven incremental and autonomous visual learning algorithm is proposed basedon the above description, with which the robot is able to actively adjust learning strategiesaccording to the visual novelty. The visual scene learning experiment is done to verify that thealgorithm has the ability of autonomous exploration and learning, actively guiding robot to learnand acquire new knowledge, and develop intelligence in an online and incremental manner.(4)Aiming to solve the problem of knowledge storage, accumulation and development, avisual novelty driven growing long-term memory (GLTM) autonomous learning algorithm isproposed to simulating the coordination mechanism of long-term memory and working memoryin human brain. The algorithm accumulates autonomously learned visual knowledge intolong-term memory and the preliminary human-like autonomous learning, growing memory andintelligence development abilities are realized. Experimental result shows that robot equippedwith GLTM is able to incrementally store and update autonomously learned knowledge. Theabilities of intelligence development, recognition, generalization and knowledge expansibilityare enhanced.(5) Aiming to solve the problem of complex nonspecific task learning, a growing long-termmemory based sensory-motor mapping autonomous learning algorithm is proposed in the firstplace. The perception and its corresponding action are learned to realize autonomouslydeveloping learning from simple perception knowledge to complex sensory-motor mappingknowledge. Then, a complex task execution method by learning from sensory-motor mapping isproposed, in which the execution ability of complex task is learned from sensor y-motor mappingsequences. The learning ability from simple task to complex task is realized and the intelligence of robot is developed. The effectiveness of this method is verified by a simulated complex taskexperiment.The contribution done in this thesis is a preliminary exploration in cognitive intelligenceand robotics, which solves the problem of traditional method to a certain extent. There is still along way to go to realize human-like intelligence.