What Is There to Be Skeptical About? About the Fantastic Success of - - PowerPoint PPT Presentation

▶
what is there to be skeptical about
SMART_READER_LITE
LIVE PREVIEW

What Is There to Be Skeptical About? About the Fantastic Success of - - PowerPoint PPT Presentation

Oct 23, 2022 269 likes •630 views

What Is There to Be Skeptical About? About the Fantastic Success of Deep Learning & The Imperative of Incorporating Priors Into Learning Oliver Brock Robotics and Biology Laboratory The Fantastic Success of Deep Learning Lots of blowout

slide-1
SLIDE 1

What Is There to Be Skeptical About?

About the Fantastic Success of Deep Learning & The Imperative of Incorporating Priors Into Learning

Oliver Brock Robotics and Biology Laboratory

slide-2
SLIDE 2

The Fantastic Success of Deep Learning

► Lots of blowout successes in CV and beyond ► End-to-end ► Pixels to Torque ► By using a general function, a gradient, and lots of data

slide-3
SLIDE 3

DNNs

Is it Science? Engineering? Protoscience? Art?

slide-4
SLIDE 4

Hennig Brand (1630–1692, depicted: 1669)

The Alchymist, in Search

  • f the Philosopher's

Stone, Discovers Phosphorus, and prays for the successful Conclusion of his

  • peration, as was the

custom of the Ancient Chymical Astrologers (1771, 1795) Joseph Wright of Derby (1734–1797)

slide-5
SLIDE 5
slide-6
SLIDE 6

NFL

David Wolpert, William Macready and many others…

slide-7
SLIDE 7

1.36 × 1050 bits/(s kg)

Bremermann’s Limit

slide-8
SLIDE 8

1075ops/s

slide-9
SLIDE 9

10103ops/s

slide-10
SLIDE 10

10200 À 10103 600 ¿ 300.000.000

slide-11
SLIDE 11

The Imperative of Incorporating Priors Into Learning

► Assumption: NFL captures a property of the problem

space

► Hypothesis 1: The problems we want to solve are

actually really simple and the NFL does not apply

► Hypothesis 2: Neural nets capture exactly the right prior

for the type of problems we want to solve

► Hypothesis 3: Hypothesis 1 & 2 together ► Hypothesis 4: We must incorporate task-specific priors

into learning 4

slide-12
SLIDE 12

I am the deep learning hammer!

slide-13
SLIDE 13

Towards Combined Tools

slide-14
SLIDE 14

Merging Two Ends of the Spectrum

Rico Jonschkowski

slide-15
SLIDE 15

Incorporate Priors Into Learning

► Representation Learning with Newton’s Laws ► Learning with Side Information

Rico Jonschkowski and Oliver Brock. Learning State Representations with Robotic Priors. Autonomous Robots. Springer US 39(3):407-428, 2015.

Direct Pattern Multi-Task Pattern

Rico Jonschkowski, Sebastian Höfer, and Oliver Brock. Patterns for Learning with Side Information. arXiv:1511.06429. 2016.

Rico Jonschkowski Sebastian Höfer Rico Jonschkowski

slide-16
SLIDE 16

Learning With Side Information: A Simple Example

f: 3 → 14 f: 5 → 30 f: 2 → 9 What is f?

Rico Jonschkowski, Sebastian Höfer, and Oliver Brock. Patterns for Learning with Side Information. arXiv:1511.06429. 2016.

slide-17
SLIDE 17

Learning With Side Information: A Simple Example

f: 3 → 9 → 14 f: 5 → 25 → 30 f: 2 → 4 → 9 What is f?

Rico Jonschkowski, Sebastian Höfer, and Oliver Brock. Patterns for Learning with Side Information. arXiv:1511.06429. 2016.

slide-18
SLIDE 18

Problem and illustrations adapted from [Thrun, Burgard, Fox: Probabilistic Robotics. MIT Press,2006.]

Toy problem: 1D hallway navigation with identical doors

Case Study: Recursive State Estimation

slide-19
SLIDE 19

Neural Histogram Filter

slide-20
SLIDE 20

Neural Histogram Filter

slide-21
SLIDE 21

Neural Histogram Filter

slide-22
SLIDE 22

Neural Histogram Filter

slide-23
SLIDE 23

Neural Histogram Filter

slide-24
SLIDE 24

Neural Histogram Filter

slide-25
SLIDE 25

Neural Histogram Filter

slide-26
SLIDE 26

Neural Histogram Filter

slide-27
SLIDE 27

Measurement Model Motion Model Motion Update Measurement Update

Learned Models (from 1600 training steps)

slide-28
SLIDE 28

Measurement Model Motion Model Motion Update Measurement Update

Learned Models (from 1600 training steps)

slide-29
SLIDE 29

Measurement Model Motion Model Motion Update Measurement Update

Learned Models (from 1600 training steps)

slide-30
SLIDE 30

NHF predictions for 32 test steps (from 1600 training steps)

slide-31
SLIDE 31

NHF predictions for 32 test steps (from 1600 training steps)

slide-32
SLIDE 32

NHF

Qualitative Comparison (models from 1600 training steps)

slide-33
SLIDE 33

NHF RNN 2x RNN LSTM 2x LSTM

Qualitative Comparison (models from 1600 training steps)

slide-34
SLIDE 34

First Results

slide-35
SLIDE 35

Conclusion

NFL

)

I am the Gaussian process saw! I am the SVM screw driver! We are all important!