• 1

Many existing approaches to bringing machine learning to blocks-based programming environments rely on embedded JavaScript or external APIs for the underlying model. As a result, some of the central aspects of the training or inference process are opaque to the user and limits the how deeply the process can be probed and understood.

In this talk, we present our recent work where we have been investigating how these concepts can be explored entirely within a blocks-based environment and demystify advanced topics such as optimization, adversarial examples, and generative adversarial training. First, we will demo our library for automatic differentiation. Then we will walk through a hands-on activity about gradient descent and invite the attendees to run it and tinker with it. Then we will demo other machine learning projects in NetsBlox including adversarial examples and building decision trees from data.