What PAIR is reading — July 2020
“What PAIR is Reading” is a reading list series compiled by PAIR team members on a rotating basis.
Meet the author
James Wexler | Software Engineer, People + AI Research
At work, I’m drawn to projects that help people peer inside of machine learning models through interactivity and visualization, with the goal of helping them understand how a particular model works, or investigate things like machine learning fairness concerns.
I lead the work on the What-If Tool and have helped with a bunch of other projects such as TCAV and the super-fun-to-play Bach Google Doodle. As a developer working on a research team, I’m involved in many different parts of the project lifecycle: prototyping, designing, reviewing, coding, testing, documenting, launching, maintaining, and advocating.
I’m also passionate about music: both making music with friends, and seeing live music (and combining machine learning and music in ridiculous ways). And nothing beats exploring the world with my wife and son.
What I’m reading
Chris Olah and others, “Zoom In: An Introduction to Circuits”, Distill
Neural networks are often referred to as black boxes because their inner workings can be inscrutable. But there has been a ton of great research on uncovering the inner workings of these networks to better understand what they learn and how they make decisions. This work uses concepts from cellular biology to zoom into each individual neuron, and the connections between them, in image detection networks, to reason about what makes them work. The end results are simple, powerful visualizations that provide real intuition into how these models recognize objects.
Joy Buolamwini, Compassion through Computation: Fighting Algorithmic Bias
In this video, Joy Buolamwini gives a performance of her powerful spoken word poem “AI, Ain’t I a Woman?” and describes her research into the failures of facial recognition systems on black women. Joy is a leading voice in the field of discovering and mitigating algorithmic bias, and the work of her, her collaborators, and other researchers in the field is having tangible, positive impact in the world. The current trend of major technology companies eliminating or restricting use of their facial recognition algorithms can be traced back to the “Gender Shades” project from Joy and Timnit Gebru.
Will Knight, “An AI Pioneer Wants His Algorithms to Understand the ‘Why’”, Wired
Much of modern machine learning can be thought of as a giant correlation machine that associates a large set of manually-collected training examples with human-provided labels for those examples, and then tries to extrapolate from those learned correlations onto unseen examples. But, if instead those machines could reason about causality (“Why do images with black circular shapes at the bottom often contain a car?”), they might be less likely to fail (or at least more understandable in exactly why and how they failed) in scenarios where they encounter situations unseen during training.
Cassie Kozyrkov, AI = “Automated Inspiration”
Cassie has the superpower of being able to distill complex statistical concepts into really entertaining and informative articles. This article is part history lesson, part introduction to the gotchas of data analytics, and part call to action.
Joy DeGruy, A Trip to the Grocery Store
Joy, through her personal story about checking out at a grocery store, does a great job of elucidating systemic racial inequity. In just three minutes, the story shines a light on white privilege and the power of speaking up instead of being a passive observer.
Dave Grohl, The Day the Live Concert Returns, The Atlantic
As the COVID-19 pandemic persists, I find myself more and more missing the cathartic experience of going to a great concert. Dave Grohl captures the magic and beauty of live music in this essay about the power of concerts and his hopes for their eventual return. Someday I hope to see you out there, with a ticket stub in your hand.
