The Code Libray Design Of General Feedforward Neural Networks

Neural networks have been applied in many fields, such as pattern recognition, image processing, stock analysis, robot control etc. Its implemention can be classified generally into two categories:software and hardware. The implemention on software features that the microprocessor or software programs which done the majority workload of logic and arithmetic, while lots of workloads done by the specific arithmetic unit in hardware.The advantage is simple, flexible, and easy to upgrade on software, but its disadvantage is that the system run slowly and inefficient. However the advantage in hardware is that it could be customed for specific object, efficient and fast, except for high cost, poor flexibility and difficulty in upgrading. The situation had been broken since the reconfigurable technology was inventioned. Programmable logic devices are the commonly used ones, and its function can be changed by reseting the internal structure,which make the implemention in hardware has the software characteristics of easy-to-upgrade and revising.In this paper, four kinds of general feedforward networks was chosen as the object of implementation in hardware。With elaborate analysis on their principles and architecture, it has been implemented their code library based on the reconfigurable platform considering the structure which has proposed to implementing the general feedforward networks. For different type and size of neural networks in practice, it can flexible implement the wanting network by select the appropriate codes from the code library under the proposed architecture. Finally, with two examples, it has been verified the feasibility and correctness of the design. The research which implemented neural network in hardware has provided an meaningful theory and support in technology...