Field programmable gate array based sigmoid function implementation using differential lookup table and second order nonlinear function
Artificial neural network (ANN) is an established artificial intelligence technique that is widely used for solving numerous problems such as classification and clustering in various fields. However, the major problem with ANN is a factor of time. ANN takes a longer time to execute a huge number of...
Saved in:
| 主要作者: | |
|---|---|
| 格式: | Thesis |
| 語言: | English |
| 出版: |
2021
|
| 主題: | |
| 在線閱讀: | http://umpir.ump.edu.my/id/eprint/38461/1/Field%20programmable%20gate%20array%20based%20sigmoid%20function%20implementation%20using%20differential%20lookup%20table%20and%20second%20order%20nonlinear%20function.ir.pdf |
| 標簽: |
添加標簽
沒有標簽, 成為第一個標記此記錄!
|
| 總結: | Artificial neural network (ANN) is an established artificial intelligence technique that is widely used for solving numerous problems such as classification and clustering in various fields. However, the major problem with ANN is a factor of time. ANN takes a longer time to execute a huge number of neurons. In order to overcome this, ANN is implemented into hardware namely field-programmable-gate-array (FPGA). However, implementing the ANN into a field-programmable gate array (FPGA) has led to a new problem related to the sigmoid function implementation. Often used as the activation function for ANN, a sigmoid function cannot be directly implemented in FPGA. Owing to its accuracy, the lookup table (LUT) has always been used to implement the sigmoid function in FPGA. In this case, obtaining the high accuracy of LUT is expensive particularly in terms of its memory requirements in FPGA. Second-order nonlinear function (SONF) is an appealing replacement for LUT due to its small memory requirement. Although there is a trade-off between accuracy and memory size. Taking the advantage of the aforementioned approaches, this thesis proposed a combination of SONF and a modified LUT namely differential lookup table (dLUT). The deviation values between SONF and sigmoid function are used to create the dLUT. SONF is used as the first step to approximate the sigmoid function. Then it is followed by adding or deducting with the value that has been stored in the dLUT as a second step as demonstrated via simulation. This combination has successfully reduced the deviation value. The reduction value is significant as compared to previous implementations such as SONF, and LUT itself. Further simulation has been carried out to evaluate the accuracy of the ANN in detecting the object in an indoor environment by using the proposed method as a sigmoid function. The result has proven that the proposed method has produced the output almost as accurately as software implementation in detecting the target in indoor positioning problems. Therefore, the proposed method can be applied in any field that demands higher processing and high accuracy in sigmoid function output |
|---|