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

Computations for the softmax function are significantly expensive when thenumber of output classes is large. In this paper, we present a novel softmaxinference speedup method, Doubly Sparse Softmax (DS-Softmax), that leveragessparse mixture of sparse experts to efficiently retrieve top-k classes.Different from most existing methods that require and approximate a fixedsoftmax, our method is learning-based and can adapt softmax weights for abetter inference speedup. In particular, our method learns a two-levelhierarchy which divides entire output class space into several partiallyoverlapping experts. Each expert is sparse and only contains a subset of outputclasses. To find top-k classes, a sparse mixture enables us to find the mostprobable expert quickly, and the sparse expert enables us to search within asmall-scale softmax. We empirically conduct evaluation on several real-worldtasks, including neural machine translation, language modeling and imageclassification, and demonstrate that significant computation reductions can beachieved at no performance loss.

Doubly Sparse: Sparse Mixture of Sparse Experts for Efficient Softmax Inference