Deep Learning as a Service What is it that users want? Wolfgang zu - - PowerPoint PPT Presentation

deep learning as a service what is it that users want
SMART_READER_LITE
LIVE PREVIEW

Deep Learning as a Service What is it that users want? Wolfgang zu - - PowerPoint PPT Presentation

Deep Learning as a Service What is it that users want? Wolfgang zu Castell castell@helmholtz-muenchen.de Helmholtz Zentrum Mnchen DEEP-Hybrid-DataCloud has received funding from the European Unions Horizon 2020 research and innovation


slide-1
SLIDE 1

DEEP-Hybrid-DataCloud has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 777435.

Deep Learning as a Service What is it that users want?

Wolfgang zu Castell castell@helmholtz-muenchen.de Helmholtz Zentrum München

slide-2
SLIDE 2

https://deep-hybrid-dat atac acloud.eu Oct 10, 2018 2/14

Do we still need science?

Chris Anderson, Wired 2008

slide-3
SLIDE 3

https://deep-hybrid-dat atac acloud.eu Oct 10, 2018 3/14

Do we still need science?

Chris Anderson, Wired 2008

slide-4
SLIDE 4

https://deep-hybrid-dat atac acloud.eu Oct 10, 2018 4/14

Artificial neural networks

Source: P . Shrivastava

slide-5
SLIDE 5

https://deep-hybrid-dat atac acloud.eu Oct 10, 2018 5/14

Artificial Intelligence Machine Learning Representation Learning

Clarifying terminology

Deep Learning

slide-6
SLIDE 6

https://deep-hybrid-dat atac acloud.eu Oct 10, 2018 6/14

Increase in number of neurons

Source: I. Goodfellow

1 Perceptron 4 Rumelhart 11 GPU accel. CNN 12 Deep Boltzmann 18 Multi-GPU CNN 20 GoogLeNet

slide-7
SLIDE 7

https://deep-hybrid-dat atac acloud.eu Oct 10, 2018 7/14

Increase in size of data sets

Source: I. Goodfellow

slide-8
SLIDE 8

https://deep-hybrid-dat atac acloud.eu Oct 10, 2018 8/14

How does deep learning work?

Source: I. Goodfellow

functional mapping of pixels

  • nto objects is very complicated

deep learning is solving this issue by breaking up the problem into series of nested simple mappings

slide-9
SLIDE 9

https://deep-hybrid-dat atac acloud.eu Oct 10, 2018 9/14

Machine learning cycle

slide-10
SLIDE 10

https://deep-hybrid-dat atac acloud.eu Oct 10, 2018 10/14

Who is the user?

domain expertise mac achine lear arning expertise technological expertise

level of knowledge being required depends on the specific use case

slide-11
SLIDE 11

https://deep-hybrid-dat atac acloud.eu Oct 10, 2018 11/14

Retinopathy use case

Source: Eulenberg et al.

Step 1: provide trained network such that

  • thers can use it for prediction

Step 2: provide trained network for new classification task (transfer learning) Step 3: completely retrain network on new dataset Step 4: completely retrain network under data privacy constraints

slide-12
SLIDE 12

https://deep-hybrid-dat atac acloud.eu Oct 10, 2018 12/14

The ecosystem of tools & libraries

T ensorFlow: speech and image recognition (Google Brain T eam) Keras as: Python NN library (Francois Challet, Google) PyT

  • rch: DL library (Facebook KI)

Caf affe: DL library (UC Berkeley) mxnet: scalable DL framework (Apache) computer vision

  • num. lin. alg.
  • sci. comp.

plotting

slide-13
SLIDE 13

https://deep-hybrid-dat atac acloud.eu Oct 10, 2018 13/14

Architecture for deep learning

slide-14
SLIDE 14

https://deep-hybrid-datacloud.eu

This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 777435.

Thank you Any Questions?