Interpretability
The unreasonable impact of machine learning techniques demonstrate that they are here to stay. That being the case, it is critically that when an important decision is made by an algorithm, it be possible for the people affected by that decision to understand how the algorithm arrived at it's conclusion. This season is a broad exploration of explainability and interpretability techniques for AI and ML.
Interpretability
Machine learning has shown a rapid expansion into every sector and industry. With increasing reliance on models and increasing stakes for the decisions of models, questions of how models actually work are becoming increasingly important to ask.
This episode includes an interview with Aaron Roth author of The Ethical Algorithm.
Wiebe van Ranst joins us to talk about a project in which specially designed printed images can fool a computer vision system, preventing it from identifying a person. Their attack targets the popular YOLO2 pre-trained image recognition model, and thus, is likely to be widely applicable.
We welcome Su Wang back to Data Skeptic to discuss the paper Distributional modeling on a diet: One-shot word learning from text only.
Enrico Bertini joins us to discuss how data visualization can be used to help make machine learning more interpretable and explainable.
Andrei Barbu joins us to discuss ObjectNet - a new kind of vision dataset.
Walt Woods joins us to discuss his paper Adversarial Explanations for Understanding Image Classification Decisions and Improved Neural Network Robustness with co-authors Jack Chen and Christof Teuscher.
We welcome back Marco Tulio Ribeiro to discuss research he has done since our original discussion on LIME.
Kyle and Linhda discuss how Shapley Values might be a good tool for determining what makes the cut for a home renovation.
Pramit Choudhary joins us to talk about the methodologies and tools used to assist with model interpretability.
Announcing Journal Club
I am pleased to announce Data Skeptic is launching a new spin-off show called "Journal Club" with similar themes but a very different format to the Data Skeptic everyone is used to.
Jessica Hullman joins us to share her expertise on data visualization and communication of data in the media. We discuss Jessica’s work on visualizing uncertainty, interviewing visualization designers on why they don't visualize uncertainty, and modeling interactions with visualizations as Bayesian updates.
Computer Vision is not Perfect
Julia Evans joins us help answer the question why do neural networks think a panda is a vulture. Kyle talks to Julia about her hands-on work fooling neural networks.
Becca Taylor joins us to discuss her work studying the impact of plastic bag bans as published in Bag Leakage: The Effect of Disposable Carryout Bag Regulations on Unregulated Bags from the Journal of Environmental Economics and Management. How does one measure the impact of these bans? Are they achieving their intended goals? Join us and find out!
Dan Elton joins us to discuss self-explaining AI. What could be better than an interpretable model? How about a model wich explains itself in a conversational way, engaging in a back and forth with the user.
What does it mean to understand a neural network? That’s the question posted on this arXiv paper. Kyle speaks with Tim Lillicrap about this and several other big questions.
Jayaraman Thiagarajan joins us to discuss the recent paper Calibrating Healthcare AI: Towards Reliable and Interpretable Deep Predictive Models.
Daniel Kang joins us to discuss the paper Testing Robustness Against Unforeseen Adversaries.
Deep neural networks are undeniably effective. They rely on such a high number of parameters, that they are appropriately described as “black boxes”.
Uri Hasson joins us this week to discuss the paper Robust-fit to Nature: An Evolutionary Perspective on Biological (and Artificial) Neural Networks.
Deb Raji joins us to discuss her recent publication Saving Face: Investigating the Ethical Concerns of Facial Recognition Auditing.
Sungsoo Ray Hong joins us to discuss the paper Human Factors in Model Interpretability: Industry Practices, Challenges, and Needs.
Erik Härkönen joins us to discuss the paper GANSpace: Discovering Interpretable GAN Controls. During the interview, Kyle makes reference to this amazing interpretable GAN controls video and it’s accompanying codebase found here. Erik mentions the GANspace collab notebook which is a rapid way to try these ideas out for yourself.
