Jan 31, 2021

#58: Partnership on AI's AI Incident Database, and lots of productized AI

By Leon Overweel (Dynamically Typed) • Issue #58 • View online

Hey everyone, welcome to Dynamically Typed #58! In today’s issue I wrote about the AI Incident Database, a new project by the Partnership on AI that’s meant to help future machine learning researchers and developers avoid repeating bad outcomes that AI-powered systems have historically caused.

Beyond that, I’ve got quite a few productized AI links today: Facebook launched a big update to their feature that helps blind and visually impaired people understand images in their newsfeed; an engineer at Dropbox explained how ML prioritization can help reduce compute costs; and a problem with the live speech translation feature Google Translate proved to be an interesting case study of nonfunctional requirements for AI-powered products. And for ML research, I covered a new entry in the Distill Circuits thread on high-low frequency detectors. (Not too much was happening in climate AI and AI art these past weeks, so I don’t have any links for those sections today.)

Productized Artificial Intelligence 🔌

The Discover app of the AI Incident Database

The Partnership on AI to Benefit People and Society (PAI) is an international coalition of organizations with the mision “to shape best practices, research, and public dialogue about AI’s benefits for people and society.” Its 100+ member organizations cover a broad range of interests and include leading AI research labs (DeepMind, OpenAI); several universities (MIT, Cornell); most big tech companies (Google, Apple, Facebook, Amazon, Microsoft); news media (NYT, BBC); and humanitarian organizations (Unicef, ACLU).

PAI recently launched a new project: the AI Incident Database. The AIID mimics the FAA’s airplane accidents database and is similarly meant to help “future researchers and developers avoid repeated bad outcomes.” It’s launching with a set of 93 incidents, including an autonomous car that killed a pedestrian, a trading algorithm that caused a flash crash, and a facial recognition system that caused an innocent person to be arrested (see DT #43). For each incident, the database includes a set of news articles that reported about it: there are over 1,000 reports in the AIID so far. It’s also open source on GitHub, at PartnershipOnAI/aiid.

Systems like this (and Amsterdam’s AI registry, for example) are a clear sign that productized AI is a quickly starting to mature as a field, and that lots of good work is being done to manage its impact. Most importantly, I hope these projects will help us have more sensible discussions about regulating AI. Benedict Evans’ essays Notes on AI Bias and Face recognition and AI ethics are excellent reads on this; he compares calls to “regulate AI” to wanting to regulate databases — it’s not the right level of abstraction, and we should be thinking about specific policies to address specific problems instead. A dataset of categorized AI incidents, managed by a broad coalition of organizations, sounds like a great step in this direction.

Quick productized AI links 🔌

👓 Facebook has launched a significantly improved version of its automatic alternative text (AAT) feature, which helps blind or visually impaired people understand the contents of images in their Facebook feed. As explained in Facebook’s tech blog post, this new version of AAT can recognize over 1,200 distinct concepts. Interestingly, the model was trained on weakly supervised data, using the hashtags on billions of public Instagram images as labels. So if you’ve ever posted a picture of your latte and tagged it #latte on Instagram, you may have had a tiny impact on this feature. The blog post also details the user research that went into improving AAT — something I think we usually don’t hear enough about (or do enough of!) in productized AI — so make sure to give it a read. (I wish I could credit the person who wrote this post, but sadly Facebook keeps these posts anonymous, which seems a bit out of character for the company.)
📑 Win Suen wrote about a machine learning system running in production at Dropbox that decides for which files previews should be rendered: Cannes: How ML saves us $1.7M a year on document previews. She goes into two design considerations for building a highly performant AI system: the cost-benefit tradeoff of ML-powered infrastructure savings (rendering fewer previews to save compute vs. hurting user experience by not having previews) and the model complexity tradeoff (prediction accuracy vs. interpretability and cost of deployment). The final model is a gradient-boosted classifier that can “predict previews up to 60 days after time of pre-warm with >70% accuracy.”
💱 Naveen Arivazhagan and Colin Cherry wrote a post for the Google AI Blog about how they solved a problem with the live speech translation feature in Google Translate: translations would frequently get updated as more of the transcribed text became available, which users found distracting. It’s a cool glimpse into all the stuff besides just model accuracy and speed that are important to get right for a successful AI-powered product, and into how engineers think about turning these nonfunctional requirements into measurable performance metics they can optimize for.

Machine Learning Research 🎛

Synthetic tuning curves: responses of six high-low frequency detectors to artificial stimuli. (Schubert et al., 2021)

🦚 Ludwig Schubert, Chelsea Voss and Chris Olah published a new entry to the Distill Circuits thread, in which they model connections in trained convolutional neural networks as logical circuits to figure out how they work; I covered what makes this research so interesting last April in Distill: Early Vision in CNNs. Using feature visualization, dataset examples, and synthetic tuning curves, this new article goes in-depth on a relatively unintuitive class of neurons: High-Low Frequency Detectors, which activate when they encounter “directional transitions from low to high spatial frequency.” In one very cool section of the article, the authors combine clusters of high- and low-frequency circuit components into two generic HF- and LF-factors, and show that they play the same roles in the implementation of high-low frequency detectors as their individual components do. As always, the article is a great weekend long read.

I’ve also collected all 75+ ML research tools previously featured in Dynamically Typed on a Notion page for quick reference. ⚡️

Thanks for reading! As usual, you can let me know what you thought of today’s issue using the buttons below or by replying to this email. If you’re new here, check out the Dynamically Typed archives or subscribe below to get a new issues in your inbox every second Sunday.

If you enjoyed this issue of Dynamically Typed, why not forward it to a friend? It’s by far the best thing you can do to help me grow this newsletter. 🕘

Did you enjoy this issue?

By Leon Overweel (Dynamically Typed)

My thoughts and links on productized artificial intelligence, machine learning technology, and AI projects for the climate crisis. Delivered to your inbox every second Sunday.

Tweet

Share

If you don't want these updates anymore, please unsubscribe here.

If you were forwarded this newsletter and you like it, you can subscribe here.