|Autors: Gatchev, G. K., Mollov, V. S.|
Title: BoolHash: A New Convolutional Algorithm for Boolean Activations
Keywords: neural networks,
Abstract: Integer algorithms, where applicable, can both decrease the memory requirements and improve the speed of the convolutional neural networks (CNN). Boolean activations can further increase the speed gain. Here, we propose a convolutional algorithm called BoolHash. It is based on precalculated inference lookup tables (PCILTs). In addition, it uses activation merging to additionally increase the inference speed. We used a CNN with INT16 input weights, INT8 filter weights and boolean activations to compare the speed of BoolHash to that of a classic weight-adder (WA) convolutional algorithm.
Вид: публикация в международен форум, публикация в реферирано издание, индексирана в Scopus