Culpability in AI failures, Fooling NNs with NNs, AI for cancer screenings, and Epsilon Greedy Multi-Armed Bandits

06/07/19 • 22 min

This week we’re diving into some deeper impacts of AI’s successes and failures- asking where responsibility lies for an algorithm’s failures, and the endless benefits of accessibility and responsibility that come with AI implemented in healthcare. We’re also taking a deep dive into Epsilon greedy multi-armed bandits and how we can more accurately describe our successes (and our failures) in AI.
When algorithms mess up, the nearest human gets the blame by Karen Hao (MIT Technology Review)
Google Trained Its AI to Predict Lung Cancer by Christine Fisher (Engadget)
No This AI Can’t Finish Your Sentence by Tiernan Ray (ZDNet)
Introduction to Multi-Armed Bandits with Applications in Digital Advertising by Dave King (SpotX)
Register for EGG NYC, Dataiku's human-centered AI conference on June 20th to hear from leaders in the AI & advanced analytics space, including WIRED, Twitter, Hinge, and more!

https://nyc.egg.dataiku.com/

https://nyc.egg.dataiku.com/

Previous Episode

Biased Data & the Perfect Answer, Multi-Armed Bandits, and the GPUs Behind Your Neural Networks

On episode two of the podcast, Triveni and Will look at how digital assistants may perpetuate biased data, how multi-armed bandits can build a top-notch recommendation system (and win over Triveni’s heart), and their interview with Mark Buckler, PhD candidate at Cornell and author of the article, “How to Make Bad Deep Learning Hardware” on why understanding hardware may be the key to building your best models yet.

Learn more about the articles referenced in this episode below:
How Digital Virtual Assistants Like Alexa Amplify Sexism by Morgan Meaker (OneZero)
The Data Nutrition Project by Kasia Chmielinski and Sara Newman
Invisible Women: Data Bias in a World Designed for Men by Caroline Criado-Perez
Your Client Engagement Program Isn’t Doing What You Think It Is by Patric Glynn and Divya Prabhakar (Stichfix)
How to Make Bad Deep Learning Hardware by Mark Buckler (Cornell)

Next Episode

The future of data according to predictions, Python 3.0, and people.

On episode 4 of Banana Data, we’re taking a look at how our data is changing. With models in the wild skewing our future data sets, the impending shift to Python 3.0, and navigating a public distrust of Machine Learning, Triveni and Will talk through how our current decisions in AI will heavily influence its future. They’ll also take a stab at explaining GANs - in English.

Learn more about the articles referenced in this episode below:
Here’s a prediction: In the future, predictions will only get worse by Alison Schrager (Quartz)
Python’s Caduceus Syndrome: What happens when a programming language grows up by Normcore Tech (Substack)
How a Feel-Good AI Story Went Wrong in Flint by Alexis C. Madrigal (The Atlantic)