Progressive Stochastic Binarization David Hartmann & Michael - - PowerPoint PPT Presentation

β–Ά
progressive stochastic binarization
SMART_READER_LITE
LIVE PREVIEW

Progressive Stochastic Binarization David Hartmann & Michael - - PowerPoint PPT Presentation

Progressive Stochastic Binarization David Hartmann & Michael Wand Institute of Computer Science, Johannes Gutenberg-University of Mainz EMC2 WORKSHOP @ NEURIPS 2019 EMC2 WORKSHOP @ NEURIPS 2019 Stochastic Floats = 2


slide-1
SLIDE 1

Progressive Stochastic Binarization

David Hartmann & Michael Wand Institute of Computer Science, Johannes Gutenberg-University of Mainz

EMC2 WORKSHOP @ NEURIPS 2019

slide-2
SLIDE 2

EMC2 WORKSHOP @ NEURIPS 2019

slide-3
SLIDE 3

π‘₯ = 𝑑 β‹… 2𝑓 β‹… 1 + πΆπ‘ž

Stochastic Floats

π‘₯ = 𝑑 β‹… 2𝑓 β‹… 𝑛

EMC2 WORKSHOP @ NEURIPS 2019

slide-4
SLIDE 4

Dynamic Control of Accuracy

EMC2 WORKSHOP @ NEURIPS 2019

Average of Results Samples of one stochastic filter

slide-5
SLIDE 5

Use with any* Model

*Not with Backbone-Batchnorms

EMC2 WORKSHOP @ NEURIPS 2019

slide-6
SLIDE 6

Computational Attention

EMC2 WORKSHOP @ NEURIPS 2019

ResNet50v2: Input

Approximation Error: blue = smaller error

Layer 8 Layer 9 (after skip) Last Conv Layer