🍊 The Juice: The Treachery of Images

Zumo Labs presents The Juice, a weekly newsletter focused on computer vision problems (and sometimes just regular problems). Get it while it’s fresh.

Michael Stewart
Zumo Labs


Week of March 8–12, 2021


“I think I finally, sort of, almost get it now.”

We’ve all experienced the challenge of explaining technical concepts to our less technically inclined friends and family members. That’s been especially true for us at Zumo Labs, as successfully explaining synthetic data is dependent on a foundational understanding of computer vision.

That’s why last week we published a super lightweight explainer on our blog. The piece was inspired by a cute (but lacking) children’s book titled “Neural Networks for Babies,” and uses René Magritte’s arguably most famous painting, “The Treachery of Images,” as a foothold.

You can read it here. And if you find it helpful — like the Zumo Labs family member who provided the opening quote — please share it with the luddites in your life. (Heck, go ahead and forward this whole email their way. It sounds like they could use it.)



“Long term, progress in AI will come from programs that just watch videos all day and learn like a baby,” says Yann LeCun, deep learning pioneer and Facebook’s chief AI scientist. LeCun doesn’t think the current standard of training on labeled data will scale. Instead, his team has developed an algorithm that uses self-supervised learning. The algorithm, SEER, was trained on a billion Instagram images, and apparently decides for itself which images look alike. Performance so far suggests that self-supervised learning is a viable option for computer vision tasks in real-world settings.

Facebook’s New AI Teaches Itself to See With Less Human Help, via Wired.

A graph from Facebook showing that engagement on content increases as it approaches the line of being disallowed by their content policy.


“64% of all extremist group joins are due to our recommendation tools,” said a slide from an internal 2016 presentation at Facebook. In other words, they had an inkling that their AI models responsible for maximizing engagement also favored controversy, misinformation, and extremism. This deep dive on Facebook’s misinformation problem profiles the other big AI lead at Facebook, Joaquin Quiñonero Candela, and explains why his team is focused on reducing bias rather than reducing the spread of disinfo. (Hint: Facebook loves growth.)

How Facebook got addicted to spreading misinformation, via MIT Technology Review.


It’s hard to keep up with developments in the smart retail space. On the heels of AiFi opening the “world’s biggest camera-operated autonomous store” in Shanghai last month, Amazon last week opened their first international checkout-free grocery store. The London location is about the size of a convenience store and operates on the same “Just Walk Out” tech that powers their other Go locations. Meanwhile, Sweden is getting in on the action with fully cashierless stores powered by Lifvs. The container-sized shops will service remote regions, and should be an easy sell considering Sweden already has one of the most cashless economies in the world.

Shops return to rural Sweden but are now staff-free, via BBC.

Amazon UK’s first checkout-free Fresh grocery store opens in London, via Engadget.


Recently OpenAI showcased an experimental system named CLIP, which was trained on huge databases of image and text pairs. The interesting thing about CLIP is that its use of multimodal neurons means it can respond to drawings and sketches even though it was never trained on them. But there are downsides. One is that that capability makes it vulnerable to what the OpenAI team is calling a “typographic attack,” where a written word overrides whatever else is in the image. The other downside is that “multimodal neurons encoded exactly the sort of biases you might expect to find when sourcing your data from the internet.”

OpenAI’s state-of-the-art machine vision AI is fooled by handwritten notes, via The Verge.


It’s a little ironic that the big noisy street sweepers we know and love burn diesel fuel and produce over 3 million metric tons of carbon emissions each year. That inconvenient truth is at least partially responsible for the Trombia Free, the first fully electric autonomous street sweeper. Citizens of the Finnish city of Espoo are about to begin seeing this mean green clean machine on their streets this month, as the company begins a commercial pilot to fully test its capabilities and train it on real world environments. If things go well, you may see the Trombia Free in other cities beginning in 2022.

Espoo launches autonomous street sweeper pilot programme, via SmartCitiesWorld.


📄 Paper of the Week

Anycost GANs for Interactive Image Synthesis and Editing

This team from CMU, MIT, and Adobe Research has designed a GAN that can generate images at different speeds and resolutions. Choosing between speed and quality is a tradeoff that changes based on the use case: sometimes you just want a really quick preview inside photoshop. The paper also has some other bells and whistles like allowing a user to edit the latent space, but nothing we haven’t seen before. Make sure to check out the code and demo.


Think The Juice was worth the squeeze? Sign up here to receive The Juice weekly.