Multirecurrent Neural Network in Discrete Form
The paper considers paradigms for constructing a multirecurrent neural network (MRNN) of Jordan and Elman. The use of these paradigms is limited by the complexity of data processing due to real arithmetic. To simplify the implementation of neurocomputations of the considered MRNN, it is proposed to...
সংরক্ষণ করুন:
| প্রধান লেখক: | , , , , , |
|---|---|
| বিন্যাস: | Статья |
| ভাষা: | English |
| প্রকাশিত: |
Springer Science and Business Media Deutschland GmbH
2024
|
| বিষয়গুলি: | |
| অনলাইন ব্যবহার করুন: | https://dspace.ncfu.ru/handle/123456789/29205 |
| ট্যাগগুলো: |
ট্যাগ যুক্ত করুন
কোনো ট্যাগ নেই, প্রথমজন হিসাবে ট্যাগ করুন!
|
| সংক্ষিপ্ত: | The paper considers paradigms for constructing a multirecurrent neural network (MRNN) of Jordan and Elman. The use of these paradigms is limited by the complexity of data processing due to real arithmetic. To simplify the implementation of neurocomputations of the considered MRNN, it is proposed to use integer arithmetic. To solve the problem, an analysis of the continuous model of the MRNN functioning was carried out, instead of which a discrete mathematical model was proposed. The advantage of using a discrete MRNN over the known one, in addition to using integer arithmetic, is also the possibility of obtaining the topology of a neural network based on the conditions of the problem being solved. In the MRNN, the structure of the neural network is formed through a complex neural network training procedure. To confirm the correct functioning of the discrete MRNN, a computer model has been developed. Analysis of the functioning of the developed computer model on the example of time series analysis indicates the correctness of the proposed discrete model of MRNN. To compare the hardware implementation of the discrete and MRNN, their models were compiled based on the simulation of firmware for FPGA microcircuits in the VHDL language. Based on the simulation, it can be concluded that the use of a discrete network can increase performance by 17% and reduce hardware costs by 3.7 times compared to the known MRNN. |
|---|