Human-in-the-Loop Interpretability Prior Isaac Lage 1 , Andrew - - PowerPoint PPT Presentation

▶
human in the loop interpretability prior
SMART_READER_LITE
LIVE PREVIEW

Human-in-the-Loop Interpretability Prior Isaac Lage 1 , Andrew - - PowerPoint PPT Presentation

Aug 31, 2023 417 likes •660 views

Human-in-the-Loop Interpretability Prior Isaac Lage 1 , Andrew Slavin Ross 1 , Been Kim 2 , Samuel J. Gershman 1 & Finale Doshi-Velez 1 1 Harvard University & 2 Google Brain Poster: Today, 10:45 AM - 12:45 PM, Room 210 & 230 AB #119

slide-1
SLIDE 1

Human-in-the-Loop Interpretability Prior

Isaac Lage1, Andrew Slavin Ross1, Been Kim2, Samuel J. Gershman1 & Finale Doshi-Velez1

1Harvard University & 2Google Brain

Poster: Today, 10:45 AM - 12:45 PM, Room 210 & 230 AB #119

slide-2
SLIDE 2

Interpretability

clipart-library.com

slide-3
SLIDE 3

Optimizing for Interpretability

Choose a Proxy for Interpretability User Study Previous Work Optimize Proxy for Interpretability

slide-4
SLIDE 4

Optimizing for Interpretability

Choose a Proxy for Interpretability User Study Previous Work Optimize Proxy for Interpretability

How to use results to choose a better proxy? Which proxy?

slide-5
SLIDE 5

Optimizing for Interpretability

Update Model User Study

Update model directly with results! No proxy!

Human-in-the-Loop Interpretability

slide-6
SLIDE 6

Interpretability Prior

Goal: Bias model to be human interpretable Bayesian Inference

slide-7
SLIDE 7

Interpretability Prior

First: Formulate Interpretability Encouraging Prior

slide-8
SLIDE 8

Optimizing for Interpretability

Choose a Proxy for Interpretability User Study Previous Work Optimize Proxy for Interpretability Which prior captures human interpretability? Can define a prior

slide-9
SLIDE 9

Optimizing for Interpretability

Update Model User Study Human-in-the-Loop Interpretability Evaluate interpretability encouraging prior

slide-10
SLIDE 10

Interpretability Prior

Then: Identify MAP Solution First: Formulate Interpretability Encouraging Prior

slide-11
SLIDE 11

Interpretability Prior

Likelihood: Easy

Evaluate computationally No users!

slide-12
SLIDE 12

Interpretability Prior

Likelihood: Easy Prior: Hard

No closed form Evaluate with user studies! Evaluate computationally No users!

slide-13
SLIDE 13

Interpretability Prior

Challenge: Approximate MAP with few evaluations of prior Prior: Hard

No closed form Evaluate with user studies!

slide-14
SLIDE 14

Simplified Cartoon of Our Approach

Step 1: Identify Diverse, High Likelihood Models

slide-15
SLIDE 15

Simplified Cartoon of Our Approach

Candidate MAP 1: Likelihood = HIGH Candidate MAP 2: Likelihood = HIGH Candidate MAP 3: Likelihood = HIGH

Step 1: Identify Diverse, High Likelihood Models

slide-16
SLIDE 16

Simplified Cartoon of Our Approach

Candidate MAP 1: Likelihood = HIGH Prior = ? Candidate MAP 2: Likelihood = HIGH Prior = ? Candidate MAP 3: Likelihood = HIGH Prior = ?

Step 1: Identify Diverse, High Likelihood Models

slide-17
SLIDE 17

Simplified Cartoon of Our Approach

Step 2: Bayesian Optimization with User Studies

Similarity Based on Explanation Features

slide-18
SLIDE 18

Simplified Cartoon of Our Approach

Step 2: Bayesian Optimization with User Studies

Similarity Based on Explanation Features User study 1: Prior = MEDIUM

slide-19
SLIDE 19

Simplified Cartoon of Our Approach

Step 2: Bayesian Optimization with User Studies

User study 1: Prior = MEDIUM Similarity Based on Explanation Features Prior Estimate: Prior = HIGH?

slide-20
SLIDE 20

Simplified Cartoon of Our Approach

Step 2: Bayesian Optimization with User Studies

User study 2: Prior = LOW Similarity Based on Explanation Features User study 1: Prior = MEDIUM

slide-21
SLIDE 21

Simplified Cartoon of Our Approach

Step 2: Bayesian Optimization with User Studies

Similarity Based on Explanation Features User study 2: Prior = LOW User study 1: Prior = MEDIUM Prior Estimate: Prior = HIGH?

slide-22
SLIDE 22

Simplified Cartoon of Our Approach

Step 2: Bayesian Optimization with User Studies

Similarity Based on Explanation Features User study 2: Prior = LOW User study 3: Prior = HIGH User study 1: Prior = MEDIUM

slide-23
SLIDE 23

Main Takeaways

Poster: Today, 10:45 AM - 12:45 PM, Room 210 & 230 AB #119

Census Dataset MORE Interpretable Number of Iterations

  • We optimize for interpretability

directly with human feedback

  • Our approach efficiently identifies

human-interpretable and predictive models

  • MAP approximations correspond

to different interpretability proxies on different datasets