Accelerator for GoogLeNet model and method thereof

Abstract

The invention discloses an accelerator for a GoogLeNet model and a method thereof. The accelerator comprises: a controller which is used for controlling control signal flow and data flow of convolution operation and full connection operation and outputting and storing a final result in a memory; a photoelectric calculation array which is used for completing matrix vector multiplication operation after convolution operation and full connection operation mapping; an analog-to-digital converter which is used for converting the current output by the photoelectric calculation array into a digital signal; an activation function unit which is used for completing bias adding and activation function operation of convolution operation and full connection operation results; and a pooling unit which is used for completing the maximum pooling operation of the result. Convolutional operation and full connection operation of the GoogLeNet model are achieved based on the photoelectric computing array, the inference process of the GoogLeNet model can be effectively accelerated, and the method has the advantages of being high in energy efficiency, free of repeated access to off-chip storage and the like.

Description

technical field [0001] The invention relates to an accelerator of a GoogLeNet model based on an optoelectronic computing array and a method thereof, belonging to the field of optoelectronic computing and the field of neural network algorithms. Background technique [0002] Neural network is the current research hotspot in the field of artificial intelligence, and the most direct way to improve the performance of neural network is to increase the depth and width of the network, where the depth refers to the number of network layers, and the width refers to the number of neurons. But this method has the following problems: (1) there are too many parameters, if the training data set is limited, it is easy to over-fit; (2) the larger the network and the more parameters, the greater the computational complexity, and it is difficult to apply; (3) ) The deeper the network, the more prone to gradient dispersion problem (the gradient goes back and disappears easily), and it is diffic...

AI Technical Summary

Problems solved by technology

But this method has the following problems: (1) There are too many parameters, and if the training data set is limited, it is easy to produce overfitting; (2) The larger the network, the more parameters, the greater the computational complexity, and it is difficult to apply; (3) ) the deeper the network, the gradient dispersion problem is prone to occur (the gradient is easy to disappear as it traverses backwards), and it is difficult to optimize the model

However, due to the large size of the GoogLeNet model, the feature maps and weight data required for convolution operations and full connection operations are large, resulting in a huge energy consumption in data transmission, and the operation speed of traditional computing systems is slow.

Images