In an earlier post, I laid out reasons that we might want to use a differentiable decision tree for reinforcement learning. In this post, I cover some experiments comparing two approaches for learning the parameters of these models: Q-Learning and Policy Gradient, showing that Q-Learning may not be a great choice if you want to apply differentiable decision trees in your own work.
Read MoreUltimately, neural networks are lots of matrix multiplication and non-linear functions in series. And following where a single number goes and how it affects the outcome of a set of matrix multiplication problems can be rather daunting. What if we could distill a neural network into a simple decision tree, and just use that to understand how the agent makes decisions? In this overview of my recent work, I cover one approach to doing exactly that!
Read MoreIn this post, I’ll explore a bit about what interpretability means to practitioners and researchers, and two ways to accomplish the goal of understanding our machine learning systems.
Read MoreIn this short write-up, I highlight three main reasons we want interpretability from our machine learning models: fairness, safety, and utility.
Read More