Autors: Gatchev, G. K., Mollov, V. S.
Title: Combining Pre-Calculated Inference Lookup Tables in Convolutional Neural Networks
Keywords: convolutional NN, pre-calculated lookup tables, BoolHash, memory usage

Abstract: In convolutional NN, algorithms that use pre-calculated lookup inference tables (PCILTs) offer a substantial improvement of the calculation speed. However, they also have a disadvantage of using more memory. A possible approach to mitgating this disadvantage is extending these algorithms with using indirection to shared unique-values PCILT(s). Our calculations and experiments show that this can reduce the memory usage, and on hardware utilizing memory caches, can avoid delays due to exceeding memory cache size.



    Int.Scientific Conference Computer Science'2020, pp. 31-35, 2020, Bulgaria,

    Вид: публикация в международен форум, публикация в реферирано издание