A Step Back: Responsible ML

discriminative model

$\rightarrow$

input

prediction

$\rightarrow$

A Step Back: Responsible ML

discriminative model

$\rightarrow$

input

prediction

$\rightarrow$

Explaining with The Shapley Value

JT, Alexandre Luster, JS (2021) "Fast Hierarchical Games for Image Explanations", TPAMI

JT*, Beepul Bharti*, Yaniv Romano, JS (2023) "SHAP-XRT: The Shapley Value Meets Conditional Independence Testing", TMLR

Explaining with The Shapley Value

Deploying these methods to inform

better practices in real-world scenarios

JT, Paul H Yi, JS (2023) "Examination-level Supervision for Deep Learning-based Intracranial Hemorrhage Detection on Head CT", Radiology: AI

A Different Scenario

generative

model

$\rightarrow$

input

samples

$\rightarrow$

For example:

- Large Language Models

- Text-to-image Diffusion Models

- Solving Inverse Problems Stochastically

A Different Scenario

generative

model

$\rightarrow$

input

samples

$\rightarrow$

?

Running Example: CT Denoising

For an observation $y$

$$y = x + \epsilon,~\epsilon \sim \mathcal{N}(0, \sigma^2\mathbb{I})$$

reconstruct $x$ with

$$F(y) \sim \mathcal{Q}_y \approx p(x \mid y)$$

$x$

$y$

$F(y)$

Is the Model Right?

Contributions: $K$-RCPS

JT, Matt Tivnan, J W Stayman, JS (2023) "How to Trust Your Diffusion Model: A Convex Optimization Approach to Conformal Risk Control", ICML

Calibrated Quantiles

Calibrated Quantiles

$x$

$y$

lower

upper

interval

Risk Controlling Prediction Sets

Risk Controlling Prediction Sets

Risk Controlling Prediction Sets

$$\downarrow$$

Using the same $\lambda$ for all pixels is

suboptimal for interval length

$$\downarrow$$

Risk Controlling Prediction Sets

$$\downarrow$$

We want to find the shortest

intervals that control risk

$K$-RCPS: High-dimensional Risk Control

scalar $\lambda \in \mathbb{R}$

$\rightarrow$

vector $\bm{\lambda} \in \mathbb{R}^d$

$\mathcal{I}_{\lambda}(y)_j = [\text{low} - \lambda, \text{up} + \lambda]$

$\rightarrow$

$\mathcal{I}_{\bm{\lambda}}(y)_j = [\text{low} - \lambda_j, \text{up} + \lambda_j]$

$\rightarrow$

$K$-RCPS: The Procedure

1. Solve

$$\tilde{\bm{\lambda}}_K = \arg\min~\sum_{k \in [K]}n_k\lambda_k~\quad\text{s.t. empirical risk} \leq \epsilon$$

2. Choose

$$\hat{\beta} = \inf\{\beta \in \mathbb{R}:~\forall \beta' \geq \beta,~\text{risk}(M\tilde{\bm{\lambda}}_K + \beta') \leq \epsilon\}$$

$K$-RCPS: Results

Thank You!