How is Enterprise AI-Driven Problem Solving Different from Chess, Go, ImageNet, most Kaggle Challenges and Industry Moonshots?
[Please review only when 90 minutes presentations are being reviewed. This is not a 20 minute talk]
What changes when we try to address enterprise problems? How does formulating a problem differently change what questions can be answered?
What happens when our model does not capture the underlying data generating process (accuracy vs. explainability vs. generalizability)? Why does capturing correlations vs establishing causative relationships matter? Does our data even have information on questions we want answered? Does distribution of our production data look like the training data in the enterprise? [No: This assumption is very often violated, in practice]
What kind of models does it take to answer "what-if this did not happen" questions?
To examine these important questions, we will explore three increasingly complex problems solved in production to explore what it takes to successfully address problems in the enterprise with a progressive move towards dynamic causal models informed by domain knowledge and data driven machine learning.
Outline/Structure of the Talk
1. New Definition of Intelligence in the Enterprise [5 min]
2. What Breaks in the Enterprise Even if I Know All the Algorithms, Applied Them in Games, Kaggle, Moonshots? [10 min]
3. Levels of Problem Formulation, starting with cognition, to changing the field itself, to asking counterfactual questions P(y|x) vs. P(Y=y | do(X=x), z) vs. P(Y=y|X = x’, Y = y’))? [15 min]
4. Three Working Examples of Problems Solved in the Enterprise at Different Levels and What is Missed in Each
a. Adverse Impact in Hiring, Compensation, Promotions and Terminations by Gender, Ethnicity, Veteran Status and Age (Over 40) [15 min]
b. Customer Tone in Investment Banking [10 min]
c. Asking What-If Questions in Semiconductor Manufacturing and Warranty [15 min]
5. Lesson Learnt: Combinations of Hybrid ML and Causal Models [10 min]
6. Questions and Discussion [10 min]
1. What are different levels of problem formulation? Why we cannot answer certain questions by setting up the problem/hypotheses in the certain way?
2. There is no such thing as "give me all your data" and we will get you some nuggets.
3. How to do careful science of the domain to solve problems in the enterprise?
People who have experienced failed AI projects in the Enterprise in multiple domains who want to genuinely solve business problems
Prerequisites for Attendees
Understanding of challenges with multiple statistical, probabilistic, classical machine learning and deep learning techniques
Have tried to solve or would like to solve at least one end-to-end large business problem which required multiple data sources, hierarchical levels of reasoning, different types of inference, domain knowledge.