Research Of Identification Problems In Biological Neural Network Systems

Biological neural network is composed of a large number of biological neurons coupled by synapses, and is found to be one of the most complex non-linear network systems. In the field of computational neuroscience, based on mathematical model of single neurons, researchers try to establish more accurate biological neural network systems by parameter identification. Therefore, from the view of calculation, they can reveal the work mechanism of both the biological neurons and neural network systems, which gives foundation of the large-scale simulation of brain function, and foreshadows into the eventual establishment of the calculation model of the brain. And as a result, how to identify varies types of parameters in biological neural network systems with measurable neural dynamics has become a research hot spot.Modern optimization algorithms rise form last century, with a characteristic of heuristic, and are mainly used to solve complex optimization problems in practical applications. One of them is real-coded genetic algorithm, which is based on mechanism of natural selection with real solution space, high precision and no bother of binary encoding. Another one is generalized extremal optimization, which is inspired by the theory of self-organize criticality, with simple implementation, fluctuation and better searching capabilities.In this paper, the identification problems in biological neural network systems are studied based on these two modern optimization algorithms, and the robustness of the methods are evaluated by adding observation noises. The main research works are as the following:(1) Identification of parameters in single biological neuron. Two objective functions are designed based on peak detection and signal similarity respectively, combined with real-coded genetic algorithm to form a novel method. Compare with previous methods, the method proposed in this paper shows robustness against noise, has no bother of binary encoding, and needs fewer measurements, which better prepares it for practical applications.(2) Identification of connection strength in biological neural networks. Similarly two objective functions are designed based on peak detection and signal similarity respectively, combined with real-coded genetic algorithm to form a novel method. The method is able to solve the hard problem of identifying connection strength, and has achieved satisfying results, with robustness to noise, and serves as a novel idea for connection strength identification.(3) Identification of comprehensive problems in biological neural networks, including both parameters and topology. The novel generalized extremal optimization is introduced to solve the problem. With adaptive improvements, the method achieves better results compared with previous literatures. Also with robustness to noises, the method shows a good prospect for practical applications.