The Pastry A.I. That Learned to Fight Cancer | Hacker News Reader

The Pastry A.I. That Learned to Fight Cancer

newyorker.com

92 pointsjsomers19 comments

Wow,

It's weird to have what seems (I repeat, seems) like pretty a big deal technically be first described in a meandering New Yorker article.

This is a more-or-less hand-tune system apparently effectively competing with deep learning systems (identifying cancer is far from easy and not solved by deep learning at this point). As the article says, by being person-created, the system has obvious advantage in both explainability and users being able to understand it's mistakes.

Of course, a broader look from experts is in order but I can't see this as a strong demonstration of an alternative paradigm. Now, the challenge is to automate the construction of systems of this sort.

And I can find little-to-no information about this, except in Japanese.

Look up "machine teaching" or "active learning". The idea of humans helping computers or even humans directly writing the rules is not new. Heueristics added on top of alpha beta pruning were how the top chess engines worked right up until alphazero.

Isn't that what constituted most 80s "AI" research [1]?

[1] https://en.wikipedia.org/wiki/Expert_system

I find it interesting that you put scare quotes around AI to speak of the dominant trend of AI research in the last 70 years (not expert systems, but manual coding of proposed intellectual mechanisms).

I am curious. Can you explain why you use scare quotes?

Because all AI is ML, obviously. All that stuff before massive GPU farms? Not real AI. (I'm not sure why this is a common perception.)

Aside from the scare quotes, you're wrong, "machine teaching" and "expert systems" are not the same thing (and neither is considered the same as pre-deep nets computer vission). In my still-shallow research, machine teaching is a conceptual thread that began years ago but is still active and interacts deep learning currents. Expert systems are still around also, they just aren't considered AI anymore.

See: https://blogs.microsoft.com/ai/machine-teaching/ and so forth.

It's a bit different from expert systems or knowledge bases IMHO. To my limited knowledge in the medical imaging field feature engineering is still very much a trusted tool (experts tuning systems based on their domain knowledge).

https://en.wikipedia.org/wiki/Feature_engineering

I like the traditional approach. Training a brain to solve a problem is all well and good, but with a BakeryScan-like system, you can see why, and change it if necessary. It's engineering, rather than statistics.

And I liked the complementarity at the end, were they used deeplearning to virtually remove the cellophane so BakeryScan could continue to identify the pastries.

Are there any good resources for learning more classical computer vision? How would someone approach a problem like this without using machine learning?

I did a small amount of work in pre-deep learning computer vision. It was only about six months exposure but as far as I could tell, the field was a mess. There weren't general methods, just a list of specific filters, approaches and so-forth that "worked in some cases" - and a lot of leaned on a knowledge of the particular setup you were modeling - optics in some instances, Fourier analysis in others and ad-hoc processes for others.

Deep learning is general purpose, doesn't require area-experts and is more or less guaranteed to mostly work if you have sufficient data describing the task at hand.

Have you read any computer vision textbooks? Szeliski's book for example is quite popular and sort-of free:

https://szeliski.org/Book/

Not specific to computer vision, but my understanding is that the two famous AI texts by Norvig -- Principles of Artificial Intelligence Programming and Artificial Intelligence: A Modern Approach -- are introductions to the "symbolic" approach to AI (as opposed to the "computational intelligence" approaches given by the modern regimes of machine learning, deep learning, and similar), and I would imagine that methods from the symbolic approach are primarily what are used here (but I'm not an expert).

PAIP was recently made available for free download by the author [1]

[1] https://github.com/norvig/paip-lisp/releases/tag/v1.0

Though PAIP is essentially all symbolic "good old-fashioned AI" (and focuses on programming rather than AI theory), AIMA never was -- that's why they emphasized "a modern approach" in the title. I guess it's reasonable though to expect AIMA's coverage of computer vision would be good background for understanding this pastry AI.

Machine learning is not a "modern" regime. The term itself was coined in the 1950's, neural networks were first described in 1938 (see Pitts & McCulloch neuron) and the current branch of research started in the 1990's, with statistical machine learnign algorithms that succeeded propositional learners like decision tree classifiers that took off in the 1980's (Ross Quinlan's ID3 algorithm was published in 1986).

AI: A Modern Approach devotes most of its considerable bulk to symbolic AI not because there was no machine learning "back then" (2003, last edition I saw) but because for most of the hisotry of AI most of the work has been on symbolic AI and statistical machine learning was a small sub-field, sometimes not even placed under AI (for example much machine vision work was published under the "pattern recognition" rubric, often not considered part of AI).

One reason for this exclusion is that much of AI research was portrayed as an investigation of the intellectual mechanisms used by the human mind, whereas statistical machine learning was (and is) more focused on narrow applications, like image classification. Although of course the connectionist paradigm has always claimed to be emulating the human brain (or trying to), this has always been more about reproducing the structure of the human brain in a different substrate, rather than elucidating how a mind arises from a brain- which is what classical AI is primarily interested in.

So if you're wondering with the most popular textbook on AI doesn't say much about how to build a hot dog classifier, that's because classical AI does not consider that to be a burning question, on which rests our scientific understanding of the human mind.

Before deep learning (ca 2013), the SOTA was using descriptors: KAZE, SIFT, SURF, ORB etc use keypoints & descriptors based matching.[1] . Other approaches relied on shape (edge), color matching using color and shape histograms comparisons e.g. HOG [2].

Plenty of approaches existed prior to deep learning, it's just that DL just blew those out of the water with its performance.

p.s. BTW I would bucket these approaches as ML too. ML >> DL

[1] https://docs.opencv.org/master/db/d27/tutorial_py_table_of_c...

[2] https://docs.opencv.org/master/d5/d33/structcv_1_1HOGDescrip...

This doesn't describe what traditional CV approach this uses, presumably descriptor based ones: A-KAZE, SIFT etc. do work well for such scenarios. So both Deep learning and traditional machine vision have their place.

Yeah, it's a fascinating article but thin on technical details, which makes it frustrating.

The local bakery had this system put in a year or so ago, and I was hopeful they'd finally use it to accept payments other than just cash... Unfortunately not.

Reply on news.ycombinator.com