Files in this item



application/pdfECE499-Sp2019-pan.pdf (448kB)Restricted to U of Illinois
(no description provided)PDF


Title:A method for faster non-unit stride convolution in deep neural networks
Author(s):Pan, Junhao
Contributor(s):Chen, Deming
Deep Neural Network
Fourier Transform
Abstract:Since computer vision and machine learning target increasingly complicated and challenging goals, the complexity of the computation models rises rapidly as the magnitude of the datasets multiplies. Deep convolutional neural networks are implemented to many realtime applications for which faster progressing time is crucial. Thus, with the rising demand for more rapid responses from data processing, there is an urgent need for further optimized convolution algorithms. For unit stride convolutions, we use FFT-based methods and Winograd algorithms, which significantly reduce the computing complexity under some specific conditions. For non-unit stride convolutions, nevertheless, we usually cannot directly apply the algorithm mentioned above but instead use conventional direct multiplications. In this thesis, we propose an algorithm which works as an extension to both FFT and Winograd algorithms to speed up convolutions with non-unit stride. The algorithm first computes the output map as if we were performing unit stride convolution and then down-samples the calculated output map to generate the final output for non-unit stride convolution. We also present a proof of the down- sampling stage of the algorithm to confirm its accuracy. Finally, we perform tests on the method under different configurations. The results confirms that the proposed method promises accelerated processing time compared to the direct-multiplying method when computing non-unit stride convolution.
Issue Date:2019-05
Date Available in IDEALS:2019-06-17

This item appears in the following Collection(s)

Item Statistics