Pastry-recognizing Japanese AI used to fight cancer
May 4, 2021 9:55 AM   Subscribe

The Pastry A.I. That Learned to Fight Cancer. An AI program that recognizes unwrapped pastries for Japanese bakeries, built before image-recognizing neural networks became widespread, turns out to be useful for recognizing cancer cells. posted by russilwvong (32 comments total) 23 users marked this as a favorite
 
Thanks, fun article! It really seems like BRAIN happened at the exact right time, which is actually sort of a scary thought, it being the case that deep learning computer vision stuff has come so far in such a short amount of time.
posted by baptismal at 10:10 AM on May 4, 2021


Wow, I clicked through to glance at it and continued reading through to the end.
posted by infini at 10:23 AM on May 4, 2021


pretty sure that title will be the peak-21st century string of words for my day, thanks for that
posted by elkevelvet at 10:41 AM on May 4, 2021 [6 favorites]


I suspect this is all part of the plotting of Big Pigeon.
Step 1: Train AI to recognize pastries
Step 2: Train AI to recognize cancer cells
Step 3: Train pigeons to recognize cancer cells
Step 4: Train pigeons to recognize pastries
Step 5: Pigeons now have direct access to pastries!
posted by subocoyne at 10:44 AM on May 4, 2021 [16 favorites]


A surgical pathologist's obsolescence is always looming.
posted by sophrontic at 10:50 AM on May 4, 2021


I really enjoyed reading this, thank you for posting.
posted by bleep at 12:48 PM on May 4, 2021


From the Tumblr summary I don't quite understand the confusion about people preferring baked goods at a bakery to be unwrapped. Every bakery I've been to in Canada is like that too, with the exception of them also carrying bags of rolls or bagels, but even there you'd have individual unwrapped ones for people that aren't buying 6 of them.
posted by any portmanteau in a storm at 12:58 PM on May 4, 2021


The desire for pastries to be unwrapped is because they tend to give off steam and oil when they first come out of the oven, and if you don't give them some air they'll turn to an inseparable pile of mush. Also it helps to be able to see what you're ordering: japanese cake shops often use a lot of individually-wrapped-in-clear-film slicing techniques.
posted by rum-soaked space hobo at 1:18 PM on May 4, 2021 [1 favorite]


When they hand me the wrapped pastry at Starbucks instead of opening it first, I feel a visceral disappointment, like my brain was happy to pretend this was a fresh pastry that will taste good and make me happy & instead were crashing into reality and it's actually a hyper processed sugar nugget for my feedbag so I can continue producing for the overlords you know?

Also I was struck by the discussion of how many varieties of pastry are available and how the markets there respond to actual consumer demand. like wow what a concept. It's not like Americans wouldn't like all those varieties too, we just all decided not to have nice things somehow.
posted by bleep at 1:52 PM on May 4, 2021 [8 favorites]


Very cool, expert systems vs neural nets...until the end, where they blend harmoniously. I wonder how the blend can be leveraged in the cancer detection realm.

Also, did anyone else read this as the happy alternative conclusion to the Cheesoid saga?
posted by TreeRooster at 2:20 PM on May 4, 2021


The desire for pastries to be unwrapped is because they tend to give off steam and oil when they first come out of the oven, and if you don't give them some air they'll turn to an inseparable pile of mush.

No, I get that people prefer their fresh baked goods unwrapped, I just didn't understand how the person that wrote the Tumblr summary would find this to be strange, or something that was unique to the Japanese.

Also, 100 different varieties feels like something that only a couple of very large bakeries would be able to handle. The bakeries I used to go to in Japan weren't too big and would maybe have 30-40? (although I would be willing to accept that I never fully pay attention to exactly how much was on offer and just lumped different things as the same, like pastries with sausage in them, there could have been 4 or 5 different kinds but they would be all the same to me) I could believe something like the Kobeya that used to be at Shijo-Kawaramachi in Kyoto might get close to 100 though.
posted by any portmanteau in a storm at 2:22 PM on May 4, 2021 [2 favorites]


...deep learning computer vision stuff has come so far in such a short amount of time.

Interestingly I took this the other way, this story is great but it really tells me how truly dumb AI and computer vision really is. These systems know NOTHING. They are monumentally stupid. Their "knowledge" is mostly statistics that identify a few salient shapes or patterns, the way something like a mollusk sees.

The fact that a bit of a nudge in another direction and a new set of labels lets it identify cells (which really do look a bit like danishes) doesn't to me say "now the computer knows something new!" it says "the computer never knew anything about pastries to begin with!"

Of course it's amazing and I love it (I cover this stuff all the time) but the more I learn about it, and the more people I talk to, the more I understand how little these things are really capable of. It's people, who put these incredibly dumb processes to work on complex problems, that give them the illusion of intelligence or capability.
posted by BlackLeotardFront at 2:30 PM on May 4, 2021 [12 favorites]


Interestingly I took this the other way, this story is great but it really tells me how truly dumb AI and computer vision really is. These systems know NOTHING. They are monumentally stupid. Their "knowledge" is mostly statistics that identify a few salient shapes or patterns, the way something like a mollusk sees.

See also: world-leading machine vision system interprets a green apple as 99.7% likely to be an iPod if you stick a post-it with the word "ipod" scrawled on it to the apple.
posted by Superilla at 2:39 PM on May 4, 2021 [9 favorites]


I mean I don't think most ML practicioners will tell you that their models "know" things or "understand" them, but also I think you're selling it a little bit short.

It's very common to take an AI model trained for one task, lop off the end of it, add a bit, and then do very minor retraining on some other task with good results. As an analogy you could think of the first parts of an image classifying model as breaking an image down into its components, essentially, recognizing "features" in images, and the latest stages of the model tag the collection of identified features and turn it into a classification.

In lots of image classification tasks, features are features, and you just need to replace and retrain the last parts. Sometimes this works better than others, like a model trained on only flowers might never be very good and classifying cars because the features it prioritized are different (think like, fuzzy and leafy vs shiny and angular)

This is especially nice since some of the best image classifiers out there took weeks or months of multi-gpu training, so being able to spend 1 day and get a classifier for a new task is nice
posted by RustyBrooks at 2:42 PM on May 4, 2021 [2 favorites]


(and re: the apple as ipod above, yes, many of these are not at all robust outside of the settings they're designed for)
posted by RustyBrooks at 2:44 PM on May 4, 2021


...Their "knowledge" is mostly statistics...

That's true of the neural nets, or "deep learning" machines. The bakery AI, if I understand correctly, is more of an expert system: lots of human knowledge (shades of bread indicating cooking times, shapes and sizes predetermined by the bakery, etc) distilled into a series of parsing algorithms.

I think the article was saying that they wrote brand-new "artisanal" detection schemes for cancer cells, not that they nudged their old program---that "nudging" kind of repurposing, in contrast, is a strength of neural net architecture; it is how the same neural net can learn to play chess or Go.

Of course the larger point about the fact that such an algorithm cannot experience the knowledge it emulates is still valid...at least as far as we know for now!
posted by TreeRooster at 2:48 PM on May 4, 2021


The fact that a bit of a nudge in another direction and a new set of labels lets it identify cells (which really do look a bit like danishes) doesn't to me say "now the computer knows something new!" it says "the computer never knew anything about pastries to begin with!"


I wouldn't say you're wrong, but I think you might have an optimistic sense of what human image-recognition is comprised of
posted by CrystalDave at 2:48 PM on May 4, 2021 [2 favorites]


No AI here.
posted by GallonOfAlan at 3:00 PM on May 4, 2021


See also: world-leading machine vision system interprets a green apple as 99.7% likely to be an iPod if you stick a post-it with the word "ipod" scrawled on it to the apple.

Ceci n'est pas une pomme.
posted by Cardinal Fang at 3:07 PM on May 4, 2021 [7 favorites]


If a computer were truly intelligent, you could say to it 'Computer, you will never taste bacon', whereupon it would self-destruct.
posted by Cardinal Fang at 3:09 PM on May 4, 2021 [3 favorites]


These systems know NOTHING. They are monumentally stupid. Their "knowledge" is mostly statistics that identify a few salient shapes or patterns, the way something like a mollusk sees.

To be fair, this is also how I know things.
posted by It's Raining Florence Henderson at 3:10 PM on May 4, 2021 [8 favorites]


Cancer cells: the devil's pastries
posted by acb at 3:56 PM on May 4, 2021 [2 favorites]


I wouldn't say you're wrong, but I think you might have an optimistic sense of what human image-recognition is comprised of

Meat!!

Yeah, I've studied the visual system enough to know it's a total mess and mad genius at the same time... love it. One of the most interesting things in the world. Of course we've got the blobs and interblobs and other stuff all recognizing visual flow, and we've designed CV algorithms after it... we get fooled by ridiculous stuff all the time ourselves of course.

Naturally I'm exaggerating and over-simplifying, and there is some really amazing stuff in the ML world. It's just funny how often I get reminded how fragile their systems of knowledge are.
posted by BlackLeotardFront at 4:16 PM on May 4, 2021


Didn't they use giant African Rats for this too?
posted by kschang at 5:24 PM on May 4, 2021


See also: world-leading machine vision system interprets a green apple as 99.7% likely to be an iPod if you stick a post-it with the word "ipod" scrawled on it to the apple.

I mean, the same thing can happen to a person if you inject them with Key 17 before undergoing a deep interrogation to try and break through their psychic barriers and nested defense personalities.
posted by FatherDagon at 11:51 PM on May 4, 2021 [1 favorite]


If we call a computer an Apple because it has an icon of an apple on it ... can a computer call an apple an iPod because it has iPod written on it?

Or am I comparing apples and oranges?
posted by romanb at 5:48 AM on May 5, 2021 [1 favorite]


For those who are curious: here’s a link to the official Bakeryscan website which has videos of it in action. More fun things on the BRAIN Co. website.
posted by romanb at 6:28 AM on May 5, 2021


I read the post title as past tense. Like there's a robot working in a bakery that used to help find cancer. Like what happened? How did he lose such a prestigious job?
posted by Stonestock Relentless at 10:57 AM on May 5, 2021 [2 favorites]


He was abusing his subordinates. Tale as old as time 😔
posted by bleep at 11:16 AM on May 5, 2021 [1 favorite]


I would have thought it was a sort of comfy semi-retirement.
posted by pykrete jungle at 2:48 PM on May 5, 2021


Or, rather, a chance encounter with a pastry sends the AI on a journey of reminiscence.
posted by pykrete jungle at 2:48 PM on May 5, 2021 [1 favorite]


Not hotdog.
posted by TrialByMedia at 7:06 PM on May 5, 2021 [2 favorites]


« Older Extra Life: A Short History of Living Longer   |   "It comes down to freedom." Newer »


This thread has been archived and is closed to new comments