English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Low-bit quantized neural networks are of great interest in practicalapplications because they significantly reduce the consumption of both memoryand computational resources. Binary neural networks are memory andcomputationally efficient as they require only one bit per weight andactivation and can be computed using Boolean logic and bit count operations.QNNs with ternary weights and activations and binary weights and ternaryactivations aim to improve recognition quality compared to BNNs whilepreserving low bit-width. However, their efficient implementation is usuallyconsidered on ASICs and FPGAs, limiting their applicability in real-life tasks.At the same time, one of the areas where efficient recognition is most indemand is recognition on mobile devices using their CPUs. However, there are noknown fast implementations of TBNs and TNN, only the daBNN library for BNNsinference. In this paper, we propose novel fast algorithms of ternary,ternary-binary, and binary matrix multiplication for mobile devices with ARMarchitecture. In our algorithms, ternary weights are represented using 2-bitencoding and binary - using one bit. It allows us to replace matrixmultiplication with Boolean logic operations that can be computed on 128-bitssimultaneously, using ARM NEON SIMD extension. The matrix multiplicationresults are accumulated in 16-bit integer registers. We also use specialreordering of values in left and right matrices. All that allows us toefficiently compute a matrix product while minimizing the number of loads andstores compared to the algorithm from daBNN. Our algorithms can be used toimplement inference of convolutional and fully connected layers of TNNs, TBNs,and BNNs. We evaluate them experimentally on ARM Cortex-A73 CPU and comparetheir inference speed to efficient implementations of full-precision, 8-bit,and 4-bit quantized matrix multiplications.

Fast matrix multiplication for binary and ternary CNNs on ARM CPU