Circuit for realizing sigmoid activation function

Abstract

The invention relates to a circuit for realizing a sigmoid activation function, and the circuit comprises a diode circuit which comprises a first diode, a second diode, and a third diode, wherein the diode circuit is configured in a manner that the sum of the first diode current and the second diode current is the third diode current; a current transmission circuit which is used for providing current for the input end of the diode circuit; an input current-voltage conversion circuit which is used for converting current at the input end of the diode circuit into input voltage and converting current at the output end of the diode circuit into output voltage; and a voltage division circuit which is used for outputting the ratio of the input voltage to the output voltage. According to the method, the sigmoid activation function in the neural network can be realized by using a hardware circuit.

Description

technical field [0001] The invention relates to the technical field of integrated circuit design, in particular to a circuit for realizing a sigmoid activation function. Background technique [0002] With the rapid development of information technology, the development of neural network is also changing with each passing day. In general, the neural network has gone through three stages, and its conception began in 1943. W.S.McCulloch and W.Pitts proposed the M-P model, which constructed a simplified mathematical model based on the structure and working mechanism of biological neurons. The neural network fell into a trough because people found that it could not handle the XOR problem; by the end of the 1980s, thanks to the proposal of the backpropagation algorithm, the research of the neural network ushered in the second spring, but due to the amount of calculation Since the 21st century, with the emergence of GPU, the large amount of calculation is no longer the limitation ...

AI Technical Summary

Problems solved by technology

With the development of the idea of ​​integrating storage and computing, scientific researchers began to put the matrix multiplication operation in the neural network into the memory, which can reduce time overhead and increase energy efficiency. For other layers in the network, people still use The disadvantages of code processing are obvious: it is impossible to implement a fully customized neural network accelerator with circuits, and the data after matrix multiplication still has to be transported back and forth through peripheral circuits, which greatly increases energy consumption

Images