Causality and Artificial Intelligence with Arni Steingrimsson

Dr. Jerry Smith welcomes you to another episode of AI Live and Unbiased to explore the breadth and depth of Artificial Intelligence and to encourage you to change the world, not just observe it!

Dr. Jerry is joined today by Arni Steingrimsson, a Data Science Machine Learning and Artificial Intelligence in the U.S. and Mexico. He is a senior-level Data Scientist, who comes from a biomedical field. Arnie and Dr. Jerry are talking today about Causality and the crucial role it plays in the AI space.

Key Takeaways:

• What is Causality? Why is it important to Artificial Intelligence?
• • Causality is what is causing the outcome; from a data perspective there are certain features that will be causal to the outcome but there is no guarantee that you can change the outcome by changing those features.
  • Defining causality is less important than knowing what is capable.
  • Granger causality is defined as a statistical dependence.
  • Judea Pearl proposes three levels of causality: Association, Intervention, and Counterfactual.
• Why it is important to actually know the cause of something?
• • People who want to be ahead and business leaders need to know how they can influence their decisions and make a change, that is why knowing the causality is crucial.
• Counterfactual Causality explains the connection between x and y, but y does not really change the possibility for x to occur or not to occur.
• • What are counterfactuals? They are a comparison of different states in the same world, but how do you quantitatively compute these two states? It is done by holding to a variable.
• Simpson’s paradox: Something observed at a high level is counter to the thing observed at a low level.
• • Simpson’s paradox is usually overlooked.
  • The study of data is an important part of the causality world.
• Using machine learning in the world of causality:
• • There are some data scientists that didn’t study causality, and they think that they can just use classical machine learning, isolating features, and feature reduction and that means using causality... but that is not the way of “changing the world”; you need to know why certain inputs changed and what caused this change.
  • A reported driver is different than a causal driver.
• The application of Evolutionary principles in the AI world:
• • The predictors are the blocks that put those inputs which are causal; this way we know the causal input to then create the machine learning model that will tell what will happen as a result of the given inputs but it does not tell us what we should set those inputs to.
  • First, we figure out what is causal and make a model for that, then once we have this model of the world, we tell people what conditions need to be set to get the best chances of achieving your outcome.
  • What kinds of tools are used for evolutionary computing? Python and their library called Deep.
• What can be done after simulation? What is next?
• • After simulation, we need to take the inputs that represent causal drivers and put them into action in the field to monitor the change.
  • If you want to improve your product you need to put programs (such as marketing and sales efforts) out and collect the data on them, how they are improving and what are the changes.

Stay Connected with AI Live and Unbiased:

Visit our website AgileThought.com

Email your thoughts or suggestions to [email protected] or Tweet @AgileThought using #AgileThoughtPodcast!

Learn more about Dr. Jerry Smith

Mentioned in this episode:

Causality: Models, Reasoning, and Inference, by Judea Pearl