CNN convolution kernel hardware design method based on cascade form

Abstract

The invention discloses a CNN convolution kernel hardware design method based on a cascade form. The method comprises the first step of defining a convolution fundamental matrix, and adopting hardware implementation to obtain a convolution fundamental module; the second step of presenting linear convolution using a fast convolution algorithm in a form of matrix multiplication, and conducting cascading on the convolution fundamental module according to the presenting mode to form a cascade convolution module; the third step of adding an addition matrix which is achieved by a summator into the cascade convolution module to achieve parallel processing, addition operation of an offset value and two-dimensional convolution sum so that the final fast convolution kernel can be obtained. According to the CNN convolution kernel hardware design method based on the cascade form, compared with a traditional two-dimensional convolution kernel, the hardware demand is lowered, and the system velocity is increased at the same time; secondly, the shortcoming that when fast convolution kernels are different in size, a hardware architecture needs to be redesigned is avoided.

Description

technical field [0001] The invention relates to the technical field of deep learning, in particular to a cascade-based CNN convolution kernel hardware design method. Background technique [0002] Over the past few years, deep learning has excelled at solving many problems such as visual recognition, speech recognition, and natural language processing. Among the different types of neural networks, convolutional neural networks are the most intensively studied. In the early days, due to the lack of training data and computing power, it was difficult to train high-performance convolutional neural networks without overfitting. [0003] There has been a plethora of state-of-the-art results in Convolutional Neural Network research recently. Since 2006, many methods have been designed to overcome the difficulty of training deep convolutional neural networks (CNN). Among them, the most famous is Krizhevsky et al. proposed a classic CNN structure - AlexNet, and made a major breakt...

AI Technical Summary

Problems solved by technology

However, doing so also increases the overall complexity of the network, greatly increasing the demand for computation

In CNN calculation, the convolution operation accounts for more than 90% of the total calculation amount. Due to its structural characteristics, the traditional von Neumann computer architecture is difficult to quickly realize the CNN operation.

Images