Flexibly Fair Representation Learning by Disentanglement Elliot - - PowerPoint PPT Presentation

▶
flexibly fair representation learning by disentanglement
SMART_READER_LITE
LIVE PREVIEW

Flexibly Fair Representation Learning by Disentanglement Elliot - - PowerPoint PPT Presentation

Jan 22, 2024 443 likes •539 views

Flexibly Fair Representation Learning by Disentanglement Elliot Creager 1 2 David Madras 1 2 J orn-Henrik Jacobsen 2 Marissa A. Weis 2 3 Kevin Swersky 4 Toniann Pitassi 1 2 Richard Zemel 1 2 June 13, 2019 1 University of Toronto 2 Vector

slide-1
SLIDE 1

Flexibly Fair Representation Learning by Disentanglement

Elliot Creager1 2 David Madras1 2 J¨

  • rn-Henrik Jacobsen2

Marissa A. Weis2 3 Kevin Swersky4 Toniann Pitassi1 2 Richard Zemel1 2 June 13, 2019

1University of Toronto 2Vector Institute 3University of T¨

ubingen

4Google Research Creager et al. 2019, arXiv:1906.02589, poster # 131 1

slide-2
SLIDE 2

Why Fair Representation Learning?

Fair Representation: [x, a] f → z

g1

→ ˆ y1 z

g2

→ ˆ y2 z

g3

→ ˆ y3 . . . Given sensitive attribute a ∈ {0, 1}, we want:

  • z ⊥ a (demographic parity) with z = f (x, a)
  • z maintains as much info about x as possible

A fair representation acts as a group parity bottleneck Current approaches are flexible w.r.t. downstream task labels (Madras et al., 2018) but inflexible w.r.t. sensitive attributes

Creager et al. 2019, arXiv:1906.02589, poster # 131 2

slide-3
SLIDE 3

Further Motivation

Subgroup discrimination

  • We would like to handle the case where a ∈ {0, 1}Na is a vector of sensitive

attributes

  • ML systems can discriminate against subgroups defined via conjunctions of

sensitive attributes (Buolamwini & Gebru, 2018)

(Kim & Mnih, 2018)

Disentangled Representation Learning

  • Each dimension of z should correspond to

no more than one semantic factor of variation (object shape, position, etc.) in the data

  • VAE variants encourage factorized

posterior q(z|x) (Higgins et al., 2017) or aggregate posterior q(z) (Kim & Mnih, 2018; Chen et al., 2018)

Creager et al. 2019, arXiv:1906.02589, poster # 131 3

slide-4
SLIDE 4

Flexibly Fair VAE

Data flow at train time (left) and test time (right) for FFVAE

Desiderata

  • z ⊥ bj ∀ j
  • bi ⊥ bj ∀ i = j
  • MutualInfo(aj, bj) is large ∀ j

Latent Code Modification

  • To achieve DP w.r.t. some ai, use

[z, b]\bi

  • To achieve DP w.r.t. conjunction
  • f binary attributes ai ∧ aj ∧ ak,

use [z, b]\{bi, bj, bk}

Creager et al. 2019, arXiv:1906.02589, poster # 131 4

slide-5
SLIDE 5

Flexibly Fair VAE

Data flow at train time (left) and test time (right) for FFVAE Learning Objective LFFVAE(p, q) = Eq(z,b|x)[log p(x|z, b) + α

  • j

log p(aj|bj)] − γDKL(q(z, b)||q(z)

  • j

q(bj)) − DKL [q(z, b|x)||p(z, b)] α encourages predictiveness in the latent code; γ encourages disentanglement

Creager et al. 2019, arXiv:1906.02589, poster # 131 5

slide-6
SLIDE 6

Results - Synthetic Data

DSpritesUnfair

Samples

Figure: With correlated factors of variation, a fair classification task is predicting Shape without discriminating against XPosition

0.0000 0.0005 0.0010 0.0015 0.0020 0.0025 DP 0.70 0.75 0.80 0.85 0.90 0.95 1.00

Accuracy

FFVAE FactorVAE CVAE

  • VAE

MLP

a = Scale

0.00 0.05 0.10 0.15 0.20 0.25 0.30 DP 0.70 0.75 0.80 0.85 0.90 0.95 1.00

Accuracy

FFVAE FactorVAE CVAE

  • VAE

MLP

a = Shape

0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 DP 0.70 0.75 0.80 0.85 0.90 0.95 1.00

Accuracy

FFVAE FactorVAE CVAE

  • VAE

MLP

a=Shape∧Scale

0.00 0.05 0.10 0.15 0.20 0.25 DP 0.70 0.75 0.80 0.85 0.90 0.95 1.00

Accuracy

FFVAE FactorVAE CVAE

  • VAE

MLP

a=Shape∨Scale Figure: Pareto-fronts showing fairness-accuracy tradeoff curves, DSpritesUnfair dataset. Optimal point is top left corner (perfect accuracy, no unfairness). y = XPosition.

∆DP | E[ˆ y = 1|a = 1] − E[ˆ y = 1|a = 0]| with ˆ y ∈ {0, 1}

Creager et al. 2019, arXiv:1906.02589, poster # 131 6

slide-7
SLIDE 7

Results - Tabular and Image Data

Communities & Crime

0.0 0.1 0.2 0.3 0.4 0.5 0.6 DP 0.65 0.70 0.75 0.80 0.85

Accuracy

FFVAE FactorVAE CVAE

  • VAE

typical success:

a = R∧P

0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14 0.16 DP 0.74 0.76 0.78 0.80 0.82 0.84

Accuracy

FFVAE FactorVAE CVAE

  • VAE

typical failure:

a = R∨B

  • Neighborhood-level population

statistics: 120 features for 1, 994 neighborhoods

  • We choose racePctBlack (R),

blackPerCap (B), and pctNotSpeakEnglWell (P) as sensitive attributes

  • Held-out label

y = violentCrimesPerCapita

Celeb-A

0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 DP 0.625 0.650 0.675 0.700 0.725 0.750 0.775 0.800 0.825

Accuracy

FFVAE

  • VAE

typical success: a = ¬ E ∧ M

0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 DP 0.625 0.650 0.675 0.700 0.725 0.750 0.775 0.800 0.825

Accuracy

FFVAE

  • VAE

typical failure: a = C ∧ M

  • Over 200, 000 images of celebrity

faces, each associated with 40 binary attributes (OvalFace, HeavyMakeup, etc.)

  • We choose Chubby (C), Eyeglasses

(E) and Male (M) as sensitive attributes

  • Held-out label

y = HeavyMakeup (H)

Creager et al. 2019, arXiv:1906.02589, poster # 131 7

slide-8
SLIDE 8

Conclusion

  • FFVAE enables flexibly fair downstream classification by

disentangling information from multiple sensitive attributes

  • Future work: extending to other group fairness definitions,

and studying robustness of disentangled/fair representation learners to distribution shift

  • Visit us at poster # 131 tonight!

Creager et al. 2019, arXiv:1906.02589, poster # 131 8