flâneur — a map of the web's best reading

GLU Variants Improve Transformer | HTML5

ar5iv.labs.arxiv.org · 3,364 words · saved by 1 readers

Gated Linear Units [Dauphin et al., 2016] consist of the component-wise product of two linear projections, one of which is first passed through a sigmoid function. Variations on GLU are possible, using different nonlinear (or even linear) functions in place of sigmoid. We test these variants in the feed-forward sublayers of the Transformer [Vaswani et al., 2017] sequence-to-sequence model, and find that some of them yield quality improvements over the typically-used ReLU or GELU activations. The Transformer [Vaswani et al., 2017] sequence-to-sequence model alternates between multi-head attention, and what it calls "position-wise feed-forward networks" (FFN). The FFN takes a vector 𝑥 (the hidden representation at a particular position in the sequence) and passes it through two learned linear transformations, (represented by the matrices 𝑊 1 and 𝑊 2 and bias vectors 𝑏 1 and 𝑏 2 ). A rectified-linear (ReLU) [Glorot et al., 2011] activation function applied between the two li

GLU Variants Improve Transformer Noam Shazeer Google noam@google.com Abstract Gated Linear Units [Dauphin et al., 2016 ] consist of the component-wise product of two linear projections, one of which is first passed through a sigmoid function. Variations on GLU are possible, using different nonlinear (or even linear) functions in place of sigmoid. We test these variants in the feed-forward sublayers of the Transformer [Vaswani et al., 2017 ] sequence-to-sequence model, and find that some of them yield quality improvements over the typically-used ReLU or GELU activations. 1 Introduction The Tran

Explore this link on the map →

related reading