\definecolor{cb1}{RGB}{76,114,176} \definecolor{cb2}{RGB}{221,132,82} \definecolor{cb3}{RGB}{85,168,104} \definecolor{cb4}{RGB}{196,78,82} \definecolor{cb5}{RGB}{129,114,179} \definecolor{cb6}{RGB}{147,120,96} \definecolor{cb7}{RGB}{218,139,195} \newcommand{\abs}[1]{\left\lvert #1 \right\rvert} \newcommand{\norm}[1]{\left\lVert #1 \right\rVert} \DeclareMathOperator*{\argmin}{argmin} \DeclareMathOperator{\bigO}{\mathcal{O}} \DeclareMathOperator{\D}{\mathcal{D}} \DeclareMathOperator*{\E}{\mathbb{E}} \DeclareMathOperator*{\Var}{Var} \DeclareMathOperator{\mean}{mean} \DeclareMathOperator{\sign}{sign} \DeclareMathOperator{\spn}{span} \newcommand{\tp}{^\mathsf{T}} \DeclareMathOperator{\tr}{tr} \newcommand{\ud}{\mathrm{d}} \DeclareMathOperator{\W}{\mathcal{W}} \DeclareMathOperator{\KL}{KL} \DeclareMathOperator{\JS}{JS} \DeclareMathOperator{\mmd}{MMD} \DeclareMathOperator{\smmd}{SMMD} \DeclareMathOperator{\hsic}{HSIC} \DeclareMathOperator{\mmdhat}{\widehat{MMD}} \DeclareMathOperator{\hsichat}{\widehat{HSIC}} \newcommand{\optmmd}[1][\Psic]{\operatorname{\mathcal{D}_\mathrm{MMD}^{#1}}} \DeclareMathOperator{\optmmdhat}{\hat{\mathcal{D}}_\mathrm{MMD}} \newcommand{\optsmmdp}{\operatorname{\mathcal{D}_\mathrm{SMMD}}} \newcommand{\optsmmd}[1][\Psic]{\operatorname{\mathcal{D}_\mathrm{SMMD}^{\SS,#1,\lambda}}} \newcommand{\ktop}{k_\mathrm{top}} \newcommand{\lip}{\mathrm{Lip}} \newcommand{\cH}{\mathcal{H}} \newcommand{\h}{\mathcal{H}} \newcommand{\R}{\mathbb{R}} \newcommand{\cP}[1]{{\color{cb1} #1}} \newcommand{\PP}{\cP{\mathbb P}} \newcommand{\pp}{\cP{p}} \newcommand{\X}{\cP{X}} \newcommand{\Xi}{\cP{X_i}} \newcommand{\Xp}{\cP{X'}} % kexpfam colors \newcommand{\nc}{{\color{#d62728}{n}}} \newcommand{\Xc}{{\color{#d62728}{X}}} \newcommand{\Mc}{{\color{#1f77b4}{M}}} \newcommand{\Yc}{{\color{#1f77b4}{Y}}} \newcommand{\mc}{{\color{#17becf}{m}}} \newcommand{\dc}{{\color{#2ca02c}{d}}} \newcommand{\Pdata}{\cP{\mathbb{P}_\mathrm{data}}} \newcommand{\cQ}[1]{{\color{cb2} #1}} \newcommand{\QQ}{\cQ{\mathbb Q}} \newcommand{\qq}{\cQ{q}} \newcommand{\qtheta}{\cQ{q_\theta}} \newcommand{\Y}{\cQ{Y}} \newcommand{\Yp}{\cQ{Y'}} \newcommand{\Yj}{\cQ{Y_j}} \newcommand{\thetac}{\cQ{\theta}} \newcommand{\vtheta}{\thetac} \newcommand{\Qtheta}{\QQ_\thetac} \newcommand{\Gtheta}{\cQ{G_\theta}} \newcommand{\cZ}[1]{{\color{cb5} #1}} \newcommand{\Z}{\cZ{Z}} \newcommand{\Zc}{\cZ{\mathcal Z}} \newcommand{\ZZ}{\cZ{\mathbb Z}} \newcommand{\cpsi}[1]{{\color{cb3} #1}} \newcommand{\psic}{\cpsi{\psi}} \newcommand{\Psic}{\cpsi{\Psi}} \newcommand{\Dpsi}{\cpsi{D_\psi}} \newcommand{\SS}{\cpsi{\mathbb{S}}} \newcommand{\Xtilde}{\cpsi{\tilde{X}}} \newcommand{\Xtildep}{\cpsi{\tilde{X}'}}

Are these datasets the same?
Learning kernels for efficient and fair two-sample tests

Danica J. Sutherland(she/her)
University of British Columbia (UBC) / Alberta Machine Intelligence Institute (Amii)
Hsiao-Yu (Fish) Tung
Heiko Strathmann
Soumyajit De
Aaditya Ramdas
Alex Smola
Arthur Gretton
Feng Liu
Wenkai Xu
Jie Lu
Guangquan Zhang
Arthur Gretton
Feng Liu
Wenkai Xu
Jie Lu
Namrata Deka

TrustML - 15 Feb 2022

(Swipe or arrow keys to move through slides; m for a menu to jump; ? to show more.)

Data drift

  • The textbook ML setting:
    • Train on i.i.d. samples from some distribution, \Xi \sim \PP
    • Training error \approx test error on \PP
    • So our model should be good on more samples from \PP
  • Really:
    • Train on “i.i.d. samples from some distribution, \Xi \sim \PP
    • Training error might vaguely correlate with test error on \PP
    • Deploy it on some distribution \QQ , might be sort of like \PP
      • and probably changes over time…

This talk

Based on samples \{ \Xi \} \sim \PP and \{ \Yj \} \sim \QQ :

  • How is \PP different from \QQ ?
  • Is \PP close enough to \QQ for our model?
  • Is \PP = \QQ ?

Two-sample testing

  • Given samples from two unknown distributions \X \sim \PP \qquad \Y \sim \QQ
  • Question: is \PP = \QQ ?
  • Hypothesis testing approach: H_0: \PP = \QQ \qquad H_1: \PP \ne \QQ
  • Reject H_0 if test statistic \hat T(\X, \Y) > c_\alpha
  • Do smokers/non-smokers get different cancers?
  • Do Brits have the same friend network types as Americans?
  • When does my laser agree with the one on Mars?
  • Are storms in the 2000s different from storms in the 1800s?
  • Does presence of this protein affect DNA binding? [MMDiff2]
  • Do these dob and birthday columns mean the same thing?
  • Does my generative model \Qtheta match \Pdata ?
  • Independence testing: is P(\X, \Y) = P(\X) P(\Y) ?

What's a hypothesis test again?

Permutation testing to find c_\alpha

Need \Pr_{H_0}\left( \hat T(\X, \Y) > c_\alpha \right) \le \alpha


c_\alpha : 1-\alpha th quantile of \left\{ \fragment[1]{\hat T(\cP{\tilde X_1}, \cQ{\tilde Y_1}), \;} \fragment[2]{\hat T(\cP{\tilde X_2}, \cQ{\tilde Y_2}), \;} \fragment[3]{\cdots} \right\}

Need a \hat T to estimate the difference between distributions, based on samples

Our choice of \hat T : the Maximum Mean Discrepancy (MMD)

This is a kernel-based distance between distributions

What's a kernel again?

  • Linear classifiers: \hat y(x) = \sign(f(x)) , f(x) = w\tp \left( x, 1 \right)
  • Use a “richer” x : f(x) = w\tp \left( x, x^2, 1\right) = w\tp \phi(x)
  • Can avoid explicit \phi(x) ; instead k(x, y) = \langle \phi(x), \phi(y) \rangle_{\cH}
  • “Kernelized” algorithms access data only through k(x, y) f(x) = \langle w, \phi(x) \rangle_\cH = \sum_{i=1}^n \alpha_i k(X_i, x)
  • \lVert f \rVert_\cH = \sqrt{\alpha\tp K \alpha} gives kernel notion of smoothness

Reproducing Kernel Hilbert Space (RKHS)

  • Ex: Gaussian RBF / exponentiated quadratic / squared exponential / … k(x, y) = \exp\left( - \frac{\norm{x - y}^2}{2 \sigma^2} \right)
  • Some functions with small \lVert f \rVert_\cH :

Maximum Mean Discrepancy (MMD)

\mmd_k(\PP, \QQ) = \sup_{\lVert f \rVert_\cH \le 1} \E_{\X \sim \PP}[f(\X)] - \E_{\Y \sim \QQ}[f(\Y)]

The sup is achieved by f(t) \propto \mathbb E_{\X \sim \PP}[f(t)] - \mathbb E_{\Y \sim \QQ}[f(t)]

\mmd^2(\PP, \QQ) = \E_{\substack{\X, \Xp \sim \PP\\\Y, \Yp \sim \QQ}}\left[ k(\X, \Xp) + k(\Y, \Yp) - 2 k(\X, \Y) \right]

\begin{align} \mmd&_k(\PP, \QQ) = \sup_{\lVert f \rVert_\cH \le 1} \E_{\X \sim \PP}[f(\X)] - \E_{\Y \sim \QQ}[f(\Y)] \\&\fragment[1]{ = \sup_{\lVert f \rVert_\cH \le 1} \E_{\X \sim \PP}[\langle f, \varphi(\X) \rangle_\cH] - \E_{\Y \sim \QQ}[\langle f, \varphi(\Y) \rangle_\cH] } \\&\fragment[2]{ = \sup_{\lVert f \rVert_\cH \le 1} \left\langle f, \E_{\X \sim \PP}[\varphi(\X)] - \E_{\Y \sim \QQ}[\varphi(\Y)] \right\rangle_\cH } \\&\fragment[3]{= %\left\lVert % \E_{\X \sim \PP}[\varphi(\X)] %- \E_{\Y \sim \QQ}[\varphi(\Y)] %\right\rVert_\cH \sup_{\lVert f \rVert_\cH \le 1} \left\langle f, \mu^k_\PP - \mu^k_\QQ \right\rangle_\cH } \fragment[4]{= \left\lVert \mu^k_\PP - \mu^k_\QQ \right\rVert_\cH } \\ \fragment[5]{\langle \mu_\PP^k, \,} & \fragment[5]{% \mu_\QQ^k \rangle_\cH = \E_{\substack{\X \sim \PP\\\Y \sim \QQ}} \langle \varphi(\X), \varphi(\Y) \rangle_\cH = \E_{\substack{\X \sim \PP\\\Y \sim \QQ}} k(\X, \Y) } \end{align}

\mmd^2(\PP, \QQ) = \E_{\substack{\X, \Xp \sim \PP\\\Y, \Yp \sim \QQ}}\left[ k(\X, \Xp) + k(\Y, \Yp) - 2 k(\X, \Y) \right]

Estimating MMD

\begin{gather} \mmd_k^2(\PP, \QQ) % = \E_{\substack{\X, \Xp \sim \PP\\\Y, \Yp \sim \QQ}}\left[ % k(\X, \Xp) % - 2 k(\X, \Y) % + k(\Y, \Yp) % \right] = \E_{\X, \Xp \sim \PP}[k(\X, \Xp)] + \E_{\Y, \Yp \sim \QQ}[k(\Y, \Yp)] - 2 \E_{\substack{\X \sim \PP\\\Y \sim \QQ}}[k(\X, \Y)] \\ \fragment[0]{ \mmdhat_k^2(\X, \Y) = \fragment[1][highlight-current-red]{\mean(K_{\X\X})} + \fragment[2][highlight-current-red]{\mean(K_{\Y\Y})} - 2 \fragment[3][highlight-current-red]{\mean(K_{\X\Y})} } \end{gather}




MMD as feature matching

\mmd_k(\PP, \QQ) = \left\lVert \E_{\X \sim \PP}[ \varphi(\X) ] - \E_{\Y \sim \QQ}[ \varphi(\Y) ] \right\rVert_{\cH}

  • \varphi : \Xc \to \cH is the feature map for k(x, y) = \langle \varphi(x), \varphi(y) \rangle
  • If k(x, y) = x\tp y , \varphi(x) = x , then
    the MMD is distance between means
  • Many kernels: infinite-dimensional \cH

MMD-based tests

  • If k is characteristic, \mmd(\PP, \QQ) = 0 iff \PP = \QQ
  • Efficient permutation testing for \mmdhat(\X, \Y)
    • H_0 : n \mmdhat^2 converges in distribution
    • H_1 : \sqrt n ( \mmdhat^2 - \mmd^2 ) asymptotically normal
  • Any characteristic kernel gives consistent test…eventually
  • Need enormous n if kernel is bad for problem

Classifier two-sample tests

  • \hat T(\X, \Y) is the accuracy of f on the test set
  • Under H_0 , classification impossible: \hat T \sim \mathrm{Binomial}(n, \frac12)
  • With k(x, y) = \frac14 f(x) f(y) where f(x) \in \{-1, 1\} ,
    get \mmdhat(\X, \Y) = \left\lvert \hat{T}(\X, \Y) - \frac12 \right\rvert

Deep learning and deep kernels

  • k(x, y) = \tfrac14 f(x) f(y) is one form of deep kernel
  • Deep models are usually of the form f(x) = w\tp \phi_\psic(x)
    • With a learned \phi_\psic(x) : \mathcal X \to \R^{D}
  • If we fix \psic , have f \in \cH_\psic with k_\psic(x, y) = \phi_\psic(x)\tp \phi_\psic(y)
    • Same idea as NNGP approximation
  • Generalize to a deep kernel: k_\psic(x, y) = \kappa\left( \phi_\psic(x), \phi_\psic(y) \right)

Normal deep learning \subset deep kernels

  • Take k_\psic(x, y) = \tfrac14 f_\psic(x) f_\psic(y) \fragment[1]{+ 1}
  • Final function in \cH_\psic will be a f_\psic(x) \fragment[1]{+ b}
  • With logistic loss: this is Platt scaling

So what?

  • This definitely does not say that deep learning is (even approximately) a kernel method
  • …despite what some people might want you to think
  • We know theoretically deep learning can learn some things faster than any kernel method [see Malach+ ICML-21 + refs]
  • But deep kernel learning ≠ traditional kernel models
    • exactly like how usual deep learning ≠ linear models

Optimizing power of MMD tests

  • Asymptotics of \mmdhat^2 give us immediately that \Pr_{H_1}\left( n \mmdhat^2 > c_\alpha \right) \approx \Phi\left( \frac{\sqrt n \mmd^2}{\sigma_{H_1}} - \frac{c_\alpha}{\sqrt n \sigma_{H_1}} \right) \mmd , \sigma_{H_1} , c_\alpha are constants: first term usually dominates
  • Pick k to maximize an estimate of \mmd^2 / \sigma_{H_1}
  • Use \mmdhat from before, get \hat\sigma_{H_1} from U-statistic theory
  • Can show uniform \mathcal O_P(n^{-\frac13}) convergence of estimator

Blobs dataset

Blobs kernels

Blobs results

Investigating a GAN on MNIST

CIFAR-10 vs CIFAR-10.1

Train on 1 000, test on 1 031, repeat 10 times. Rejection rates:

Ablation vs classifier-based tests

Cross-entropyMax power
High- d Gauss. mix.0.470.590.290.640.66
MNIST vs GAN0.650.710.800.941.00


  • What if you don't have much data for your testing problem?
  • Need enough data to pick a good kernel
  • Also need enough test data to actually detect the difference
  • Best split depends on best kernel's quality / how hard to find
    • Don't know that ahead of time; can't try more than one


  • One idea: what if we have related problems?
  • Similar setup to meta-learning:(from Wei+ 2018)

Meta-testing for CIFAR-10 vs CIFAR-10.1

  • CIFAR-10 has 60,000 images, but CIFAR-10.1 only has 2,031
  • Where do we get related data from?
  • One option: set up tasks to distinguish classes of CIFAR-10 (airplane vs automobile, airplane vs bird, ...)

One approach (MAML-like)

A_\theta is, e.g., 5 steps of gradient descent

we learn the initialization, maybe step size, etc

This works, but not as well as we'd hoped…
Initialization might work okay on everything, not really adapt

Another approach: Meta-MKL

Inspired by classic multiple kernel learning

Only need to learn linear combination \beta_i on test task:
much easier

Theoretical analysis for Meta-MKL

  • Same big-O dependence on test task size 😐
  • But multiplier is much better:
    based on number of meta-training tasks, not on network size
  • Coarse analysis: assumes one meta-tasks is “related” enough
    • We compete with picking the single best related kernel
    • Haven't analyzed meaningfully combining related kernels (yet!)

Results on CIFAR-10.1


  • Sometimes we know ahead of time that there are differences that we don't care about
    • In the MNIST GAN criticism, initial attempt just picked out that the GAN outputs numbers that aren't one of the 256 values MNIST has
  • Can we find a kernel that can distinguish \PP^t from \QQ^t ,
    but can't distinguish \PP^s from \QQ^s ?
  • Also useful for fair representation learning
    • e.g. can distinguish “creditworthy” vs not,
      can't distinguish by race

High on one power, low on another

Choose k with \min_k \rho_k^s - \rho_k^t

  • First idea: \displaystyle \rho = \frac{(\mmd)^2}{\sigma_{H_1}}
    • No good: doesn't balance power appropriately
  • Second idea: \displaystyle \rho = \Phi\left( \frac{\sqrt{n} (\mmd)^2 - c_\alpha}{\sigma_{H_1}} \right)
    • Can estimate c_\alpha inside the optimization
    • Better, but tends to “stall out” in minimizing \rho_k^s

Block estimator [Zaremba+ NeurIPS-13]

  • Use previous \mmdhat on b blocks, each of size B
  • Final estimator: average of each block's estimate
    • Each block has previous asymptotics
    • Central limit theorem across blocks
  • Power is \displaystyle \rho = \Phi\left( \sqrt{b B} \frac{\mmd^2}{\sigma_{H_1}^2} - \Phi^{-1}(1 - \alpha) \right)


  • Choose k as \min_k \rho_k^s - \rho_k^t
    • \rho is the power of a test with b blocks of size B
    • We don't actually use a block estimator computationally
    • b , B have nothing to do with minibatch size
  • Representation learning: \min_\phi \max_\kappa \rho^s_{\kappa \circ \phi} - \rho^t_{\kappa \circ \phi}
    • Deep kernel is [\kappa \circ \phi](x, y) = \kappa(\phi(x), \phi(y))
    • \kappa could be deep itself, with adversarial optimization
    • For now, just Gaussians with different lengthscales



\PP^t : \QQ^t :
\PP^s : \QQ^s :

Multiple targets / sensitive attributes

\displaystyle \max_k \frac{1}{\lvert \mathcal T \rvert} \sum_{t \in \mathcal T} \rho^t_k - \frac{1}{\lvert \mathcal S \rvert} \sum_{s \in \mathcal S} \rho^s_k

power on minibatches

Remaining challenges

  • MMD-B-Fair:
    • When s and t are very correlated
    • For attributes with many values (use HSIC?)
  • Meta-testing: more powerful approaches, better analysis
  • When \PP \ne \QQ , can we tell how they're different?
    • Methods so far: low- d , and/or points w/ large critic value
  • Avoid the need for data splitting (selective inference)

A good takeaway

Combining a deep architecture with a kernel machine that takes the higher-level learned representation as input can be quite powerful.

— Y. Bengio & Y. LeCun (2007), “Scaling Learning Algorithms towards AI