AI Simplified - Episode 12 - Ali El-Sharif on Explainable Artificial Intelligence
0:00 Introduction
3:56 Explainable AI
5:11 AI’s Promise
5:50 Is this a Tree?
7:23 X-ray images & when AI fails to meet its promise
10:08 Supervised learning needs more attention
11:36 High quality training data sets = costly
12:28 Cross-industry data trusts
13:09 Constructing training data sets
14:46 Data trusts in research community
16:30 Throwing computing power towards data sets
17:31 Fooling ML algorithms
18:13 ML algorithms vulnerable to attacks
19:19 Bias in ML algorithms
20:14 Amazon AI Recruiting Tool
21:08 Explainability is the law
22:11 What is Explainable AI?
23:02 Explainable AI definition by DARPA
23:25 Explainable AI Target Audience
24:31 ML Explanations Benefits
24:48 Explainability leads to adoption
25:41 David Deutsch - traits of good explanations
28:12 General relativity vs. quantum mechanics
29:53 The law & corporate responsibility
32:10 Biggest black box is the human brain
32:30 Explanation is a model of the truth
33:00 Why is explainability so difficult?
33:54 Decision Trees
34:27 Interpretable model explanations don’t scale
34:55 Use Interpretable Models - If You Can
35:34 People fall in love with complex solutions
37:57 Complexity is sometimes inevitable; GPT-3 NLP Model
38:33 Why is interpretability difficult?
39:36 White box vs. black box models
41:05 Local explanation & single predictions
43:59 Post-Hoc Explanations
44:43 Explainability Options
45:19 Local Interpretable Model-Agnostic Explanations (LIME)
48:47 Explaining a prediction with LIME
49:57 LIME - Image Classifier Explanation
50:39 LIME - Text Classifier Explanation
51:16 Pros & Cons of LIME method
53:25 Overall thoughts on LIME
53:41 Foundation for Best Practices in ML
54:46 Attraction to Explainable AI research & road ahead
59:50 AI models that have conversations with humans
1:00:50 Label requirements for explainability models
1:04:00 Need for high quality data sets & collaborative models
1:07:25 Canada’s leadership in AI
1:09:40 Diversity in Canada
1:11:15 Advice for AI practitioners
1:13:30 Closing Thoughts