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

Compressing large neural networks is an important step for their deploymentin resource-constrained computational platforms. In this context, vectorquantization is an appealing framework that expresses multiple parameters usinga single code, and has recently achieved state-of-the-art network compressionon a range of core vision and natural language processing tasks. Key to thesuccess of vector quantization is deciding which parameter groups should becompressed together. Previous work has relied on heuristics that group thespatial dimension of individual convolutional filters, but a general solutionremains unaddressed. This is desirable for pointwise convolutions (whichdominate modern architectures), linear layers (which have no notion of spatialdimension), and convolutions (when more than one filter is compressed to thesame codeword). In this paper we make the observation that the weights of twoadjacent layers can be permuted while expressing the same function. We thenestablish a connection to rate-distortion theory and search for permutationsthat result in networks that are easier to compress. Finally, we rely on anannealed quantization algorithm to better compress the network and achievehigher final accuracy. We show results on image classification, objectdetection, and segmentation, reducing the gap with the uncompressed model by 40to 70% with respect to the current state of the art.

Permute, Quantize, and Fine-tune: Efficient Compression of Neural Networks