To obtain access to full text of journal and articles you must register!
- Article name
- Optimization of the convolutional operations for usage of them in convolutional neural networks, which are implemented on the FPGA basis
- Authors
- Kushchenko A. S., , andrew.kushchenko@gmail.com, Institute of Computer Technologies and Information Security, South Federal University; JSC "Scientific Design Bureau of Computing Systems", Taganrog, Rostov region, Russia
Makarevich O. B., , mak@tsure.ru, Institute of Computer Technologies and Information Security, South Federal University, Taganrog, Rostov region, Russia
Polovko I. Yu., , i.y.polovko@gmail.com, South Federal University, Institute of Computer Technologies and Information Security, Taganrog, Rostov region, Russia
- Keywords
- neural network convolution / FPGA / image processing / artificial intelligence
- Year
- 2020 Issue 2 Pages 59 - 62
- Code EDN
- Code DOI
- Abstract
- This article discusses a classical implementation of convolutional filters, which are a part of machine learning algorithms. The advantages of FPGAs over GPU on the objective of image processing by convolutional neural networks was given. The method was proposed for a replacement of a multiplication operation with a bitwise operation xor to reduce necessary resources on the chip for a calculation of the neural network. Replacing the operation of multiplication with the bitwise operation xor is not equivalent, nonetheless the usage of Hamming distance has proved similarity of images. Thereby is proved that it is possible to replace the multiplication operation in a convolutional filter and the neural network still will be able to highlight the key features in the image, which are necessary for the work of the network. In the article were given theoretical calculations of the possible number of computational elements after replacement of the multiplication operation with the bitwise operation xor, by using the example of a FPGA Stratix chip. Intel.
- Text
- BUY for read the full text of article
- Buy
- 500.00 rub
