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/

Zendy Plus

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

Zendy Tools

Zendy Open

Modern CNN are typically based on floating point linear algebra basedimplementations. Recently, reduced precision NN have been gaining popularity asthey require significantly less memory and computational resources compared tofloating point. This is particularly important in power constrained computeenvironments. However, in many cases a reduction in precision comes at a smallcost to the accuracy of the resultant network. In this work, we investigate theaccuracy-throughput trade-off for various parameter precision applied todifferent types of NN models. We firstly propose a quantization trainingstrategy that allows reduced precision NN inference with a lower memoryfootprint and competitive model accuracy. Then, we quantitatively formulate therelationship between data representation and hardware efficiency. Ourexperiments finally provide insightful observation. For example, one of ourtests show 32-bit floating point is more hardware efficient than 1-bitparameters to achieve 99% MNIST accuracy. In general, 2-bit and 4-bit fixedpoint parameters show better hardware trade-off on small-scale datasets likeMNIST and CIFAR-10 while 4-bit provide the best trade-off in large-scale taskslike AlexNet on ImageNet dataset within our tested problem domain.

Accuracy to Throughput Trade-offs for Reduced Precision Neural Networks on Reconfigurable Logic