New open source tools to facilitate the implementation of binarized neural networks on FPGAs

Neural network technology originated in the 1950s and 1960s, when it was called a perceptron, with an input layer, an output layer and an implicit layer. The input feature vector reaches the output layer through the hidden layer transformation, and the classification result is obtained in the output layer. The promoter of the early perceptron is Ronsenblatt. Later, it developed into a multi-layer perceptron, while the multi-layer perceptron was freed from the constraints of the early discrete transfer function, and the back-propagation BP algorithm invented by Werbos was used in the training algorithm. This is the neural network NN of the current constant, and currently exists. The most common neural networks are: ANN, RNN, and CNN. CNN is a multi-layer neural network that is good at dealing with related machine learning problems of images, especially large images. It can successfully carry out image recognition problems with huge data volume through a series of methods, and finally enable them to be trained.

GUNNESS open source tool

A new open source tool called GUNNESS helps users easily implement Binary Neural Networks (BNNs) on Zynq SoC chips and Zynq UltraScale+ MPSoC chips through the SDSoC development environment. GUINNESS was developed based on GUI tools, and the internal implementation uses a deep learning framework to train a binary CNN. This part of the content is a comprehensive introduction to this year's IEEE international parallel and distributed processing workshop (the paper is entitled "on-chip Memory Based binarized Convolutional Deep Neural Network Applying Batch Normalization Free Technique on an FPGA"), in the paper, authors Haruyoshi Yonekawa and Hiroki Nakahara describe a system they implemented: they implemented a binary CNN logic system for running the VGG-16 benchmark on the Xilinx ZCU102 Eval kit, where the ZCU102 kit actually It is built on the Zynq UltraScale+ MPSoC chip. Later in the FPL2017 of Ghent, Belgium, the author Nakahara introduced the GUINNESS tool again.

According to the paper published in the IEEE, the comparison of CNN implemented on Zynq and running CNN on ARM Cortex-A57 processor speeds up by 136.8 times and increases power efficiency by 44.7 times. Compared to running the same CNN on Nvidia Maxwell GPUs, Byn based on Zynq speeds up 4.9 times faster and consumes 3.8 times more power.

However, for us, the most gratifying thing is that the entire GUINNESS tool can be accessed on Github (https://github.com/HirokiNakahara/GUINNESS).

Figure: Xilinx ZCU102 Zynq UltraScale+ MPSoC Eval Kit

The current concept of comparative fire is no more than machine learning, deep learning, and artificial intelligence. The realization of these technologies is inseparable from the training of neural networks. It can be said that the hotspots of current technology are non-neural networks. However, the neural network algorithm is often complicated, and the software implementation speed often cannot meet the demand. The dedicated chip design has a single function and high cost. However, when implemented by FPGA, it not only avoids the single-purpose and high-cost input, but also obtains the operation speed desired by the user. , killing two birds with one stone. It is also believed that in the future, FPGAs will have more room for research and implementation of neural networks.