1 of 17

transformation importance

with applications to cosmology

singh*, ha*, lanusse, boehm, liu & yu

2 of 17

input

prediction

3 of 17

MassiveNuS simulations (liu et al. 2017)

4 of 17

attribution methods

shap (lundberg & lee 2017)

contextual decomposition (murdoch et al. 2018; singh et al. 2019)

lime (ribeiro et al. 2016)

integrated gradients (sundarajan et al. 2017)

5 of 17

TRIM(s)

T -1

x’

f’(s)

f(x)

x

s

T

x - x’

6 of 17

TRIM (CD) score

Central scale (angular multipole ℓ)

7 of 17

TRIM (CD) score

Central scale (angular multipole ℓ)

8 of 17

9 of 17

fake news classification

mean importance (cd)

10 of 17

audio classification

mean importance �(integrated gradients)

11 of 17

mnist

12 of 17

simulations

X

Y

training

importance

cd

deeplift

shap

ig

0.4

3.6

4.0

4.2

error (%)

13 of 17

next steps: learning a transformation with desirable properties

cd importance

relevant part

residuals

T

14 of 17

appendix

15 of 17

prior knowledge: should be pretty flat

A_s

16 of 17

first layer filters

filters from ribli et al. 2019

17 of 17

cosmology

​

  • add realistic noise levels
  • try on Gaussian maps

posthoc interp

​

  • parametrized wavelets
  • sparse importance scores
  • some kind of bottleneck…
  • further investigating frequency domain

model-based interp

​

  • convolutional sparse coding
  • scattering transform model
  • IRM model

next steps