In this talk, Coco Sack talks about current explainability techniques in the field.
Presented on 08/21/2019
**These slides are from a talk given at Rsqrd AI. Learn more at rsqrdai.org**
Rsqrd AI: A Survey of The Current Ecosystem of Explainability Techniques
1. DECIDING IN THE DARK:A Survey of The Current
Ecosystem of Explainability
Techniques
By Coco Sack 1
2. WHO AM I?
AI2 Incubator Intern
Yale freshman (Computer Science & Psychology major)
Child of AI Boom
As of recent, an “Explainability Expert”
2
4. WHO CARES ABOUT EXPLAINABILITY?
DATA-SCIENTISTS
To debug...
POLITICIANS
GDPR:
“a right to
explanation”
To protect the public...
EXECUTIVES
76%
said lack of
transparency
was seriously
impeding
adoption
To sell...
CONSUMERS
To trust...
4
9. DATA ANALYSIS
◂ Data must be complete, unique, credible, accurate, consistent, and
unbiased
◂ Errors or duplicates can undermine the model’s performance
◂ Outdated bias in the training data can re-entrench discrimination.
You are saying, 'Here's the data, figure out what the behavior is.'
That is an inherently fuzzy and statistical approach. The real
challenge of deep learning is that it's not modeling, necessarily, the
world around it. It's modeling the data it's getting. And that modeling
often includes bias and problematic correlations.”
-- Sheldon Fernandez, CEO of DarwinAI
9
17. OTHER EXPLAINABILITY RESOURCES
General Surveys of Techniques:
A Survey of Methods For Explaining
Black Box Models (2018)
Visual Analytics in Deep Learning
(2018)
Explaining Explanations (2019)
Peeking Inside the Black Box (2018)
Books and Presentations:
Interpretable Machine Learning (by
Christopher Molnar, 2019)
XAI (by Dave Gunning, 2017)
Explaining Explanations (2019)
Unique Technical Papers:
Generative Synthesis (Wong et al.)
Golden Eye++ (Hendricks et al.)
Grad-CAM (Selvaraju et al.)
DeepLift (Shrikumar et al.)
T-CAV (Kim et al.)
Ethical/Political Reports:
“Computer Says No” (by Ian Sample)
“Why We Need to Open the Black Box”
(by AJ Abdallat)
“The Importance of Interpretable
Machine learning” (by DJ Sakar)
17
Editor's Notes
INTRINSICALLY INTERPRETABLE: the most straightforward way to achieve interpretability is to design an algorithm or model that is intrinsically interpretable meaning it is naturally human interpretable due to its simple and intuitive structures.
DEEP-EXPLANATION: altering deep learning models to create secondary systems that are trained to generate explanations of the first.
OUTPUT ANALYSIS: Finally output-focused methods can analyze an opaque machine learning model after it is trained based on its outputs. These techniques are “post-hoc,” meaning they are applied after training a black-box model making them more general, flexible, and applicable. Some of these techniques are specifically designed for certain kinds of models while others are truly agnostic and can be applied to any machine learning model.