Deep Learning and Explainable AI

Previously, we learned about many types of machine learning such as supervised machine learning, clustering, time-series analysis, and NLP. Finally, we're ready to discuss two of the hottest topics in machine learning: Deep Learning and Explainable AI.

What is Deep Learning?

So what is deep learning?

Deep Learning, also sometimes called as "neural networks" or "neural nets", is a special type of machine learning that can solve more complex problems. It requires much, much more data than traditional machine learning. It is best used in cases where inputs are less structured, such as large amounts of text or images.

Explainable AI

One of the main drawbacks of Deep Learning is a lack of explainability. Although Deep Learning can make very accurate predictions, it's not always clear why the model is making a specific prediction.

Methods that allow us to understand the factors that lead to each prediction are also know as "Explainable AI", where "AI" stands for "Artificial Intelligence".

Case: Explainable AI

Let's examine a typical problem in Explainable AI. Suppose we are investigating customer cancellations using a traditional machine learning model. Our trained model can tell us two things.

  1. Prediction: What a customer is likely to do
  1. Explanation: Why a customer is likely to do it

First, it can predict whether or not a given customer is likely to churn. Second, it can tell us which features were important in making this decision. This is "explainable" part.

This additional "explainability" can provide important insights. For instance, we might learn that certain demographics are much more likely to cancel their subscriptions. Our Marketing and Customer Support teams can now use this insight to change their outreach strategies and address this deficiency.

Prediction: Which letter is this likely to be?

Contrast that example with a typical Deep Learning problem. Suppose we want to recognize hand-written letters. We don't really care why a particular image was classified as an "a", as long as the predictions are highly accurate.

Deep learning is a perfect solution to this problem because we don't care about explainability and we probably have a large, image-based set of training data.