David Touretzky

Dave Touretzky headshot

Carnegie Mellon University

Dr. Touretzky’s research focuses on teaching artificial intelligence in K-12. He is the founder and chair of AI4K12.org, which is developing national guidelines for K-12 AI education. To make AI concepts accessible to K-12 students, Dr. Touretzky develops online tools and demonstrations that allow students to do hands-on experiments with AI systems. His work in AI-CARING has focused on three topics: visualizing knowledge graphs, visualizing transformer neural networks, and understanding the behavior of linear threshold neurons.
Knowledge graphs are used to represent knowledge in a form computers can easily reason with. Major AI companies such as Google and Microsoft have developed proprietary knowledge graph implementations, not directly accessible to the public. Open source knowledge graphs exist, but have limited scope and can be difficult to work with. Dr. Touretzky’s project is a small demonstration knowledge graph with a carefully crafted visual interface that allows non-experts to explore the representation without getting lost in a profusion of links. The assistive agents that AI-CARING is developing will likely make use of some type of knowledge graph to represent information about patients, medical conditions, household products, etc. The knowledge graph visualization techniques developed in Dr. Touretzky’s research can potentially help caregivers and the general public understand how these agents structure their knowledge.
In the second project, Dr. Touretzky is developing a tool for exploring transformer neural networks, a type of neural network architecture used by large language models (LLMs) such as BERT and ChatGPT. The tool, called BERT-insight, provides a visualization of MobileBERT, a scaled-down version of BERT that can run in the web browser on an ordinary laptop. The tool shows how inputs are tokenized, how tokens are translated to embedding vectors, how attention heads process these sequences, and how an extractive question answering module derives answers from the model’s outputs. Work is now underway to try to characterize the contributions made by specific attention heads.
In the third project, Dr. Touretzky is creating a tool to help students understand linear threshold units, the building blocks of neural networks. This tool, called Neuron Sandbox, presents a series of problems of increasing difficulty that a linear threshold unit can be used to solve. Students must determine the weights and/or threshold value required for a solution. The tool provides extensive visual scaffolding to help students consider all the input cases and what result is produced for each one. Starting from two-input AND and OR functions, problems become more complex in several ways: adding more inputs, requiring different weights on different inputs, requiring mixtures of positive and negative weights, requiring a negative threshold, and accepting non-binary inputs. The tool includes an authoring mode that allows teachers to create their own customized problems for their students to solve. Current work is investigating automatic generation of hints when a student struggles to solve a problem.