SAM: The Sensitivity of Attribution Methods to Hyperparameters (CVPR 2020) - Dr. Chirag

Attribution methods can provide powerful insights into the reasons for a classifier’s decision. Perturbation-based explanation methods often measure the contribution of an input feature to an image classifier's outputs by heuristically removing it via e.g., blurring, adding noise, or graying out, which often produce unrealistic, out-of-samples. Additionally, a key desideratum of an explanation method is its robustness to input hyperparameters which are often randomly set or empirically tuned.
High sensitivity to arbitrary hyperparameter choices does not only impede reproducibility but also questions the correctness of an explanation and impairs the trust of end-users. To this extent, integrating generative models for removing input features can ameliorate existing problems by:
(1) generating more plausible counterfactual samples under the true data distribution;
(2) being more accurate according to three metrics: object localization, deletion, and saliency metrics; and
(3) being more robust to hyperparameter changes.

The talk is based on the paper:
SAM: The Sensitivity of Attribution Methods to Hyperparameters (CVPR 2020)

00:00 Intro
02:15 Motivation 1: The Urban Legend Tank No Tank
03:15 Motivation 2: Out-of-distribution examples
07:16 Attribution maps as explanations
12:49 Saliency maps
17:34 Saliency maps may NOT be too noisy!
19:34 Smoothed gradients can be misinterpreted
20:34 Insensitivity noise
23:21 Many attribution maps are sensitive to hyperparameters
28:58 Attribution maps are more robust under robust classifiers
35:58 Sensitivity of an individual attribution map
40:58 Some hyperparameters are more detrimental
43:53 Conclusions and Discussion: 1. Gradient images for robust classifiers are smooth 2. Smoothing gradients may cause misinterpretation 3. Many attribution methods are sensitive to hyper-parameters 4. For robust classifiers, attribution maps are more robust
51:55 Explaining image classifiers by removing input features using generative models
56:15 Sliding Patch
57:06 LIME
59:36 LIME samples - Jigsaw puzzle
01:03:36 Learn a blur mask
01:05:28 Problems with heuristic removals
01:09:45 Idea: Marginalize over plausible counterfactuals
01:12:01 Results: More accurate object localization
01:17:32 More realistic counterfactuals
01:20:32 Are G-methods more robust to hyperparameter changes?
01:23:23 Discussion

[Chapters were auto-generated using our proprietary software - contact us if you are interested in access to the software]

Presenter BIO:

Bio: Chirag Agarwal is a postdoctoral research fellow at Harvard University and completed his Ph.D. in electrical and computer engineering from the University of Illinois at Chicago under the joint guidance of Dr. Dan Schonfeld and Dr. Anh Nguyen. Chirag's research has primarily focused on three pillars of performance, robustness, and explanations, which he believes are necessary for deploying machine learning models safely to practical applications. His current works, primarily, focus on different aspects of Trustworthy ML like fairness, robustness, and explainability.

-------------------------
Find us at:

[ddai]

Would love to hear any comment or input you have. Did you find the content interesting? Is there a place for improvement? Please leave any feedback you have here in a comment :)