Testing Semantic Importance
via Betting

Jeremias Sulam

 Interpretability in Modern AI
Foundations, Methods,

and Emerging Directions


UCSD 2026

 

"The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin. [...] Seeking an improvement that makes a difference in the shorter term, researchers seek to leverage their human knowledge of the domain, but the only thing that matters in the long run is the leveraging of computation. [...]
We want AI agents that can discover like we can, not which contain what we have discovered."

The Bitter Lesson, Rich Sutton 2019

  • Sensitivity or Gradient-based perturbations

  • Shapley coefficients

  • Variational formulations

  • Counterfactual & causal explanations

LIME [Ribeiro et al, '16], CAM [Zhou et al, '16], Grad-CAM [Selvaraju et al, '17]

Shap [Lundberg & Lee, '17], ...

RDE [Macdonald et al, '19], ...

[Sani et al, 2020] [Singla et al '19],..

Interpretability in Image Classification

  • Sensitivity or Gradient-based perturbations

  • Shapley coefficients

  • Variational formulations

  • Counterfactual & causal explanations

LIME [Ribeiro et al, '16], CAM [Zhou et al, '16], Grad-CAM [Selvaraju et al, '17]

Shap [Lundberg & Lee, '17], ...

RDE [Macdonald et al, '19], ...

[Sani et al, 2020] [Singla et al '19],..

  • Adebayo et al, Sanity checks for saliency maps, 2018

  • Ghorbani et al, Interpretation of neural networks is fragile, 2019

  • Shah et al, Do input gradients highlight discriminative features? 2021

Interpretability in Image Classification

  • Sensitivity or Gradient-based perturbations

  • Shapley coefficients

  • Variational formulations

  • Counterfactual & causal explanations

LIME [Ribeiro et al, '16], CAM [Zhou et al, '16], Grad-CAM [Selvaraju et al, '17]

Shap [Lundberg & Lee, '17], ...

RDE [Macdonald et al, '19], ...

[Sani et al, 2020] [Singla et al '19],..

Post-hoc Interpretability Methods

Interpretable by

construction

  • Adebayo et al, Sanity checks for saliency maps, 2018

  • Ghorbani et al, Interpretation of neural networks is fragile, 2019

  • Shah et al, Do input gradients highlight discriminative features? 2021

Interpretability in Image Classification

  • Sensitivity or Gradient-based perturbations

  • Shapley coefficients

  • Variational formulations

  • Counterfactual & causal explanations

LIME [Ribeiro et al, '16], CAM [Zhou et al, '16], Grad-CAM [Selvaraju et al, '17]

Shap [Lundberg & Lee, '17], ...

RDE [Macdonald et al, '19], ...

[Sani et al, 2020] [Singla et al '19],..

Post-hoc Interpretability Methods

Interpretable by

construction

  • Adebayo et al, Sanity checks for saliency maps, 2018

  • Ghorbani et al, Interpretation of neural networks is fragile, 2019

  • Shah et al, Do input gradients highlight discriminative features? 2021

Interpretability in Image Classification

Is the piano important for \(\hat Y = \text{cat}\)?

How can we explain black-box predictors with semantic features?

Is the piano important for \(\hat Y = \text{cat}\), given that there is a cute mammal in the image?

Semantic Interpretability of classifiers

Is the presence of \(\color{Blue}\texttt{edema}\) important for \(\hat Y = \text{lung opacity}\)?

How can we explain black-box predictors with semantic features?

Is the presence of \(\color{magenta}\texttt{devices}\) important for \(\hat Y = \texttt{lung opacity}\), given that there is \(\color{blue}\texttt{edema}\) in the image?

lung opacity
cardiomegaly
fracture
no findding

Semantic Interpretability of classifiers

Semantic Interpretability of classifiers

Embeddings

 \(H = f(X) \in \mathbb R^d\)

Concept Bank:

    \(C = [c_1, c_2, \dots, c_m] \in \mathbb R^{d\times m}\)

Semantic Interpretability of classifiers

Embeddings

 \(H = f(X) \in \mathbb R^d\)

Concept Bank:

    \(C = [c_1, c_2, \dots, c_m] \in \mathbb R^{d\times m}\)

Semantic Interpretability of classifiers

Embeddings

 \(H = f(X) \in \mathbb R^d\)

Semantics
\(Z = C^\top H \in \mathbb R^m\)

Concept Activation Vectors

(Kim et al, 2018)

\(c_\text{cute}\)

Semantic Interpretability of classifiers

Vision-language models

(CLIP, BLIP, etc... )

Concept Bank:

    \(C = [c_1, c_2, \dots, c_m] \in \mathbb R^{d\times m}\)

Semantic Interpretability of classifiers

[Bhalla et al, "Splice", 2024]

Concept Bottleneck Models (CMBs)

[Koh et al '20, Yang et al '23, Yuan et al '22 ]

  • Need to engineer a (large) concept bank

  • Performance hit w.r.t. original predictor

\(\tilde{Y} = \hat w^\top Z\)

\(\hat w_j\) is the importance of the \(j^{th}\) concept

Desiderata:   Semantic explanations with

  • Fixed original predictor (post-hoc)

  • Global and local importance notions

  • Testing for any concepts (no need for large concept banks)

  • Precise testing with guarantees (Type 1 error/FDR control)

Precise notions of importance

Global Feature Importance

\(H_{0,j} : \hat{Y} \perp\!\!\!\perp Z_j \)

cuteness
cat

Precise notions of importance

mammal, piano, whiskers, ...

Global Feature Importance

\(H_{0,j} : \hat{Y} \perp\!\!\!\perp Z_j \)

cuteness

Global Conditional Importance

\(H_{0,j} : \hat{Y} \perp\!\!\!\perp Z_j | Z_{-j}\)

cuteness
cat
cat

Precise notions of importance

Global Conditional Importance

\(H_{0,j} : \hat{Y} \perp\!\!\!\perp Z_j | Z_{-j}\)

Testing

Procedure

(at level \(\alpha\))

Reject


Fail to Reject

( \(Z_j:\) important )

Precise notions of importance

Global Conditional Importance

\(H_{0,j} : \hat{Y} \perp\!\!\!\perp Z_j | Z_{-j}\)

Testing

Procedure

(at level \(\alpha\))

[Candes et al, 2018]

\(\rightarrow\) Conditional Randomization Test (CRT)

requiring \(Z_j \sim P_{Z_j|Z_{-j}}\)

Correct testing:    \(\mathbb P [\text{rejecting} | H_{0,j}:\text{true}]\leq \alpha\)

Reject


Fail to Reject

( \(Z_j:\) important )

Precise notions of importance

Global Conditional Importance

\(H_{0,j} : \hat{Y} \perp\!\!\!\perp Z_j | Z_{-j}\)

Reject


Fail to Reject

Testing

Procedure

(at level \(\alpha\))

[Candes et al, 2018]

\(\rightarrow\) Conditional Randomization Test (CRT)

requiring \(Z_j \sim P_{Z_j|Z_{-j}}\)

\(\quad \hat{Y}|(Z_j,Z_{-j}) \overset{d}{=} \hat{Y}|Z_{-j}\)

Correct testing:    \(\mathbb P [\text{rejecting} | H_{0,j}:\text{true}]\leq \alpha\)

( \(Z_j:\) important )

Local Conditional Importance

"\(\hat{Y}(x) \perp\!\!\!\perp {Z_j}(x) | Z_{-j}(x)\)"

semantic \(j\) in \(x\)

all other semantics \(\{-j\}\) in \(x\)

Precise notions of importance

Local Conditional Importance

Semantically-conditioned features:      \(\tilde H_S \sim P_{H|Z_S = z_S} \)

\[H^{j,S}_0:~ \text{classifier}({\tilde H_{S \cup \{j\}}}) \overset{d}{=} \text{classifier}(\tilde H_S)\]

features with semantic \(j\) in \(x\)

features with semantics \(\{-j\}\) in \(x\)

we can sample from these

Precise notions of importance

"The classifier (its distribution) does not change if we condition

on concepts \(S\) vs on concepts \(S\cup\{j\} \)"

eXplanation Randomization

Test (XRT)
[Teneggi et al, 2023]

\(\hat{Y}_\text{gas pump}\)

\(Z_S\cup Z_{j}\)

\(Z_{S}\)

\(Z_j=\)

Local Conditional Importance

\(\tilde{Z}_S = [z_\text{text}, z_\text{old}, Z_\text{dispenser}, Z_\text{trumpet}, Z_\text{fire}, \dots ] \)

\(S\)

\(\tilde{Z}_{S\cup j} = [z_\text{text}, z_\text{old}, z_\text{dispenser}, Z_\text{trumpet}, Z_\text{Fire}, \dots ] \)

\(S\)

\(j\)

Precise notions of semantic importance

\[H^{j,S}_0:~ \text{classifier}({\tilde H_{S \cup \{j\}}}) \overset{d}{=} \text{classifier}(\tilde H_S)\]

  \(\tilde H_S \sim P_{H|Z_S = z_S} \)

\(\hat{Y}_\text{gas pump}\)

\(\hat{Y}_\text{gas pump}\)

\(Z_S\cup Z_{j}\)

\(Z_{S}\)

\(Z_S\cup Z_{j}\)

\(Z_{S}\)

Local Conditional Importance

\(Z_j=\)

\(Z_j=\)

\(\tilde{Z}_S = [z_\text{text}, z_\text{old}, Z_\text{dispenser}, Z_\text{trumpet}, Z_\text{fire}, \dots ] \)

\(\tilde{Z}_{S\cup j} = [z_\text{text}, z_\text{old}, Z_\text{dispenser}, z_\text{trumpet}, Z_\text{Fire}, \dots ] \)

\(S\)

\(S\)

\(j\)

Precise notions of semantic importance

\[H^{j,S}_0:~ \text{classifier}({\tilde H_{S \cup \{j\}}}) \overset{d}{=} \text{classifier}(\tilde H_S)\]

  \(\tilde H_S \sim P_{H|Z_S = z_S} \)

Testing by betting 

\(H^G_{0,j} : \hat{Y} \perp\!\!\!\perp Z_j     \iff      P_{\hat{Y},Z_j} = P_{\hat{Y}} \times P_{Z_j}\)

Testing importance via two-sample tests

\(H^{GC}_{0,j} : \hat{Y} \perp\!\!\!\perp Z_j | Z_{-j}     \iff      P_{\hat{Y}Z_jZ_{-j}} = P_{\hat{Y}\tilde{Z}_j{Z_{-j}}}\)

\(\tilde{Z_j} \sim P_{Z_j|Z_{-j}}\)

[Candes et al, 2018; Shaer et al, 2023]

[Teneggi et al, 2023]

\[H^{j,S}_0:~ \text{classifier}({\tilde H_{S \cup \{j\}}}) \overset{d}{=} \text{classifier}(\tilde H_S), \qquad \tilde H_S \sim P_{H|Z_S = z_S} \]

Testing by betting

Goal: Test a null hypothesis \(H_0\) at significance level \(\alpha\)

Standard testing by p-values

Collect data, then test, and reject if \(p \leq \alpha\)

Online testing by e-values

Any-time valid inference, monitor online and reject when \(e\geq 1/\alpha\)

[Shaer et al. 2023, Shekhar and Ramdas 2023, Podkopaev et al 2023]

Testing by betting via SKIT (Podkopaev et al., 2023)

Online testing by e-values

Any-time valid inference, track and reject when \(e\geq 1/\alpha\)

  • Consider a wealth process
    \(K_0 = 1;\)
    \(\text{for}~ t = 1, \dots \\ \quad K_t = K_{t-1}(1+\kappa_t v_t)\)

Fair game (test martingale): \(~~\mathbb E_{H_0}[\kappa_t | \text{Everything seen}_{t-1}] = 0\)

\(v_t \in (0,1):\) betting fraction

\(\kappa_t \in [-1,1]\) payoff

[Grünwald 2019, Shafer 2021, Shaer et al. 2023, Shekhar and Ramdas 2023. Podkopaev et al., 2023]

\(\mathbb P_{H_0}[\exists t \in \mathbb N: K_t \leq 1/\alpha]\leq \alpha\)

Goal: Test a null hypothesis \(H_0\) at significance level \(\alpha\)

Testing by betting via SKIT (Podkopaev et al., 2023)

Online testing by e-values

Any-time valid inference, track and reject when \(e\geq 1/\alpha\)

  • Consider a wealth process
    \(K_0 = 1;\)
    \(\text{for}~ t = 1, \dots \\ \quad K_t = K_{t-1}(1+\kappa_t v_t)\)

Fair game (test martingale): \(~~\mathbb E_{H_0}[\kappa_t | \text{Everything seen}_{t-1}] = 0\)

\(v_t \in (0,1):\) betting fraction

\(\kappa_t \in [-1,1]\) payoff

[Grünwald 2019, Shafer 2021, Shaer et al. 2023, Shekhar and Ramdas 2023. Podkopaev et al., 2023]

\(\mathbb P_{H_0}[\exists t \in \mathbb N: K_t \leq 1/\alpha]\leq \alpha\)

Goal: Test a null hypothesis \(H_0\) at significance level \(\alpha\)

\(H_0: ~ P = Q\)

\(\kappa_t = \text{tahn}({\color{teal}\rho(X_t)} - {\color{teal}\rho(Y_t)})\)

Iteratively:

\( K_t = K_{t-1}(1+\kappa_t v_t)\)

\(\color{teal}{X_t \sim P; ~ Y_t\sim Q}\)

Example: 

     Data efficient 

     Rank induced by rejection time

rejection time

rejection rate

Important Semantic Concepts
(Reject \(H_0\))

Unimportant Semantic Concepts

(fail to reject \(H_0\))

Results:     Imagenette

Type 1 error control

   False discovery rate control

\(\alpha = 0.05\)

Results:     Imagenette

Results:     Imagenette

Results:     CUB dataset

Important Semantic Concepts
(Reject \(H_0\))

Unimportant Semantic Concepts

(Fail to reject)

 

rejection time

rejection rate

0.0

1.0

CheXpert: validating BiomedVLP 

What concepts does BiomedVLP find important to predict                                 ? 

lung opacity

Results:     Imagenette

Semantic comparison of vision-language models

Summary

  • Notions of importance that are rigorously testable
  • Local conditional tests for semantic importance
  • Any-time-valid inference for efficient testing and ranking
Jacopo Teneggi
JHU

Teneggi & S., Testing Semantic Importance via Betting, Neurips (2024).

Appendix

Semantic Interpretability of classifiers

Concept Bank:     \(C = [c_1, c_2, \dots, c_m] \in \mathbb R^{d\times m}\)

Vision-language models

(training)

[Radford et al, 2021]