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

Quantizing a Deep Neural Network (DNN) model to be used on a customaccelerator with efficient fixed-point hardware implementations, requiressatisfying many stringent hardware-friendly quantization constraints to trainthe model. We evaluate the two main classes of hardware-friendly quantizationmethods in the context of weight quantization: the traditional Mean SquaredQuantization Error (MSQE)-based methods and the more recent gradient-basedmethods. We study the two methods on MobileNetV1 and MobileNetV2 using multipleempirical metrics to identify the sources of performance differences betweenthe two classes, namely, sensitivity to outliers and convergence instability ofthe quantizer scaling factor. Using those insights, we propose varioustechniques to improve the performance of both quantization methods - they fixthe optimization instability issues present in the MSQE-based methods duringquantization of MobileNet models and allow us to improve validation performanceof the gradient-based methods by 4.0% and 3.3% for MobileNetV1 and MobileNetV2on ImageNet respectively.

A Closer Look at Hardware-Friendly Weight Quantization