1 of 19

What Uncertainties do we need in Bayesian Deep Learning for Computer Vision?��A. Kendall and Y. Gal�NeurIPS, 2017

Alpay Ozkan

2 of 19

Contents

​

  • Motivation
  • Concepts
  • Key Ideas
  • Results

2

METU CENG 501 Deep Learning - Paper presentation

3 of 19

Motivation

  • Probabilistic Deep Learning
  • Why do we need this?
  • Why not assign uncertainty to predictions?

3

METU CENG 501 Deep Learning - Paper presentation

4 of 19

Concepts

​

  • Bayesian Approach

​

  • Aleatoric vs Epistemic

​

  • Homoscedastic vs Heteroscedastic

4

METU CENG 501 Deep Learning - Paper presentation

5 of 19

Bayesian Neural Network (BNN)

5

METU CENG 501 Deep Learning - Paper presentation

6 of 19

Bayesian Neural Network (BNN)

6

METU CENG 501 Deep Learning - Paper presentation

7 of 19

Bayesian Deep Learning Framework

  •  

7

METU CENG 501 Deep Learning - Paper presentation

8 of 19

Likelihood and Posterior

  •  

8

METU CENG 501 Deep Learning - Paper presentation

9 of 19

Objective

  • Dropout ~ variational bayesian inference
  • Min Objective

​

​

  • For regression

​

​

  • For classification
                  • MC integration

9

METU CENG 501 Deep Learning - Paper presentation

10 of 19

Epistemic and Aleatoric Uncertainties

  • Predictive Variance (Epistemic)

​

​

  • Predictive Mean

​

​

  • Heteroscedastic Aleatoric Uncertainty

10

METU CENG 501 Deep Learning - Paper presentation

11 of 19

Combining Uncertainties: Aleatoric + Epistemic

Model output

​

Min Objective

Log Variance (practical)

Predictive Uncertainty

11

METU CENG 501 Deep Learning - Paper presentation

12 of 19

Attenuation

  • New loss function

​

  • Learned Loss Attenuation

​

  • Roboustness to noisy data

12

METU CENG 501 Deep Learning - Paper presentation

13 of 19

Heteroscedastic Uncertainty in Classification

  • Heterosced. Classif. NN

​

  • Probability vector

​

  • Network outputs with W params

​

  • Expected log-likelihood

​

  • Stochastic Loss

​

​

13

pixel

Sample t

Class c’

METU CENG 501 Deep Learning - Paper presentation

14 of 19

Experiments

  • Pixel-wise depth regression
  • Semantic segmentation

​

  • Roboustness of loss attenuation
  • Datasets:
    • CamVid
    • Make3D
    • NYUv2

14

METU CENG 501 Deep Learning - Paper presentation

15 of 19

Semantic Segmentation Experiments

​

Road Scene Understanding Indoor Scene Segmentation

15

METU CENG 501 Deep Learning - Paper presentation

16 of 19

Depth Regression Experiments

16

METU CENG 501 Deep Learning - Paper presentation

17 of 19

Observations

17

METU CENG 501 Deep Learning - Paper presentation

18 of 19

What do we capture?

18

METU CENG 501 Deep Learning - Paper presentation

19 of 19

References

  • [1] A. Kendall and Y. Gal, “What uncertainties do we need in Bayesian deep learning for computer vision?,” arXiv [cs.CV], 2017.
  • [2] J. Vincent, “Google ‘fixed’ its racist algorithm by removing gorillas from its image-labeling tech,” The Verge, 12-Jan-2018. [Online]. Available: https://www.theverge.com/2018/1/12/16882408/google-racist-gorillas-photo-recognition-algorithm-ai. [Accessed: 13-Jan-2022].
  • [3] Sky, “Tesla driver in first self-drive fatal crash,” Sky, 01-Jul-2016. [Online]. Available: https://news.sky.com/story/tesla-driver-in-first-self-drive-fatal-crash-10330121. [Accessed: 13-Jan-2022].
  • [4] TwinEd Productions, “Bayesian neural network | deep learning,” 02-Apr-2021. [Online]. Available: https://www.youtube.com/watch?v=OVne8jDKGUI. [Accessed: 16-Jan-2022].
  • [5] “Maximum likelihood estimation of Gaussian parameters,” Github.io, 18-Aug-2017. [Online]. Available: http://jrmeyer.github.io/machinelearning/2017/08/18/mle.html. [Accessed: 16-Jan-2022].
  • [6] Y. Gal and Z. Ghahramani, “Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,” arXiv [stat.ML], 2015.

​

19

METU CENG 501 Deep Learning - Paper presentation