Conditional Independence Testing using Adversarial Neural Networks - - PowerPoint PPT Presentation

conditional independence testing using adversarial neural
SMART_READER_LITE
LIVE PREVIEW

Conditional Independence Testing using Adversarial Neural Networks - - PowerPoint PPT Presentation

Conditional Independence Testing using Adversarial Neural Networks Alexis Bellot Mihaela van der Schaar From two-sample testing to independence testing X ~ Y ~ Two sample problem: Can we say whether = ? Independence


slide-1
SLIDE 1

Conditional Independence Testing using Adversarial Neural Networks

Alexis Bellot Mihaela van der Schaar

slide-2
SLIDE 2

From two-sample testing to independence testing

Two sample problem:

Can we say whether 𝑸 = 𝑹 ? X ~ 𝑸 Y ~ 𝑹

Independence problem:

Can we say whether 𝒀, 𝒁 ⊥ (𝟏, 𝟐) ?

slide-3
SLIDE 3

From independence to conditional independence

Two sample problem:

Can we say whether 𝑸 = 𝑹 ? X ~ 𝑸 Y ~ 𝑹

Independence problem:

Can we say whether 𝒀 ⊥ 𝒁 ?

Conditional Independence problem:

Can we say whether 𝒀 ⊥ 𝒁 | 𝒂 ?

slide-4
SLIDE 4

?

Other observed gene mutations Genes of interest Disease

Can we find genes directly associated with disease?

slide-5
SLIDE 5

?

Z X Y

Many questions fit into this formalism

What data should I collect for my prediction problem ? Is it likely that interventions will affect the variable of interest ? How do we learn a Bayesian network

  • ver our data ?

Important concepts of statistics (sufficiency, ancillarity …) can be regarded as expressions

  • f conditional independence

Is our prediction rule invariant to changes in the environment ? Is our prediction rule fair ? Are sensitive attributes influential ?

slide-6
SLIDE 6

?

Z X Y

𝒀 ⊥ 𝒁 | 𝒂 iff 𝒒 𝒁 𝒀, 𝒂 = 𝒒 𝒁 𝒂)

The intuition

  • If we had access to 𝒒 𝒁 𝒂)
  • Samples from this distribution breaks

the direct dependency 𝒀 − 𝒁

  • A comparison of the dependencies

between synthetic and observed data should not reveal any differences under the null.

slide-7
SLIDE 7
  • We develop a modified GAN to sample

from 𝒒 𝒁 𝒂) with high power

➢ Better performance in high dimensions

Why you should come see our poster

  • Provably valid testing

➢ No assumptions on data distribution. ➢ Non-asymptotic error bounds.

slide-8
SLIDE 8

Join us for more details shortly East Exhibition Hall B + C #75