摘要
FPGA能够充分发挥卷积神经网络的并行特性,并在小尺寸、低功耗的条件下,实现卷积神经网络的运算,是人工智能研究和发展的新方向。其中,Softmax层函数是神经网络的输出层函数,主要用于神经网络的最后一层。首先简要介绍Softmax层函数,分析几种实现函数的方案,然后采用分段拟合的方法在MATLAB上对Softmax层函数进行逼近,对数据进行量化和分析,在FPGA平台用硬件描述语言实现Softmax层函数,并通过Vivado进行仿真,结果表明误差可以控制在较小数量级。
The FPGA can give full play to the convolution of the neural network parallel features, and under the condition of small size, low power con-sumption, realize convolutional neural network arithmetic, is the new direction of artificial intelligence research and development. The Soft-max layer function is the output layer function of the neural network, which is used primarily for the last layer of the neural network. Brieflyintroduces the Softmax layer function principle, analyzes the several functions for realizing scheme, then uses piecewise fitting method inthe MATLAB function to approximate the Softmax layer, to quantify the data and analysis, the FPGA platform, uses hardware descriptionlanguage to realize the Softmax layer function and through Vivado simulation, the results show that error can be controlled in smaller ordersof magnitude.