An investigation of multiplication error tolerances in CNN and SIFT

Chanon Khongprasongsiri; Watcharapan Suwansantisuk; Pinit Kumhom

doi:10.1117/12.2521564

22 March 2019 An investigation of multiplication error tolerances in CNN and SIFT

Chanon Khongprasongsiri, Watcharapan Suwansantisuk, Pinit Kumhom

Proceedings Volume 11049, International Workshop on Advanced Image Technology (IWAIT) 2019; 110493L (2019) https://doi.org/10.1117/12.2521564
Event: 2019 Joint International Workshop on Advanced Image Technology (IWAIT) and International Forum on Medical Imaging in Asia (IFMIA), 2019, Singapore, Singapore

Abstract

A computer vision computation requires high number of multiplications causing a bottleneck. Based on the work of Zhenhong Liu, the multiplications in these algorithms do not always require high precision provided by the processors. As a result, we can reduce computation redundancy by means of multiplication approximation. Following this approach, in this paper, we investigate two major algorithms namely convolutional neural network (CNN) and scale-invariant features transform (SIFT) to find their error tolerances due to multiplication approximation. A multiplication approximation is done by injecting a random value to each of precise multiplication value. The INRIA and OXFORD datasets were used in the SIFT algorithm analysis while the CIFAR-10 and MNIST datasets were applied for the CNN experiments. The results showed that SIFT can withstand only small percents of multiplication approximation while CNN can tolerate over 30% of multiplication approximation.

Citation Download Citation

Chanon Khongprasongsiri, Watcharapan Suwansantisuk, and Pinit Kumhom "An investigation of multiplication error tolerances in CNN and SIFT", Proc. SPIE 11049, International Workshop on Advanced Image Technology (IWAIT) 2019, 110493L (22 March 2019); https://doi.org/10.1117/12.2521564

ACCESS THE FULL ARTICLE

INSTITUTIONAL
Select your institution to access the SPIE Digital Library.

SELECT YOUR INSTITUTION

PERSONAL
Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.

PERSONAL SIGN IN

No SPIE Account? Create one

PURCHASE THIS CONTENT

SUBSCRIBE TO DIGITAL LIBRARY

50 downloads per 1-year subscription

Members: $195

Non-members: $335 ADD TO CART

25 downloads per 1 - year subscription

Members: $145

Non-members: $250 ADD TO CART

PURCHASE SINGLE ARTICLE

Includes PDF, HTML & Video, when available