13 ms·

Funny because I do have the same feeling these days: that ML is a hoax. Even funnier: I do have a master's degree in ML.

timr 6y ago
It isn't a hoax, but the OP is exactly right: if it were more usable, people would see it for what it is, and not for what the silly media narrative makes it sounds like. As long as your technology is only usable by a high priesthood, you can make it look like magic.
- nerdponx 6y ago
  Why do people feel entitled to "usable" ML at all? In the last 5 years, we have made incomprehensibly huge improvements in power and usability. It's an active field, and improvements are still coming at a steady pace. We have already revolutionized search, natural language processing & machine translation, image/audio/video processing, robotics, game AI, and advertising (for better or worse). And on top of all this, we have significantly reduced the "time to first useful model", and we have significantly lowered the math and programming requirements for building and implementing models. And now we have transfer learning, which lets any old Joe Schmo benefit from massive computing power and datasets to build small on-device models that blow away SOTA accuracy from even a few years ago. Oh, and the ML tooling ecosystem has become a substantial source of innovation in programming language design, "developer UX", and "data ops". What the fuck more do you want? The people who seem the most upset that ML isn't magic seem to be the most confused about what ML even is and does.
  - timr 6y ago
    > Why do people feel entitled to "usable" ML at all? Yeah, people are annoying, with their demands to use software themselves. It would be much easier for everyone if computers were controlled by an elite group of engineers who could hide the complexity from the rest of us. Perhaps they could wear labcoats. > What the fuck more do you want? If I knew the answer to that, life would be a lot simpler.
    - dang 6y ago
      > Yeah, people are annoying, with their demands to use software themselves. It would be much easier for everyone if computers were controlled by an elite group of engineers who could hide the complexity from the rest of us. Perhaps they could wear labcoats. That's clearly not what nerdponx meant. Can you please stick to the site guidelines? "Please respond to the strongest plausible interpretation of what someone says, not a weaker one that's easier to criticize. Assume good faith." https://news.ycombinator.com/newsguidelines.html https://news.ycombinator.com/newsguidelines.html > If I knew the answer to that, life would be a lot simpler. That was a nice de-escalation.
      - timr 6y ago
        > That's clearly not what nerdponx meant. I disagree. I made a straightforward interpretation of what was written. Given what came after the part I quoted, you really have to stretch to interpret it differently. The OP underscored the same point using profanity. I made a response that was clearly sardonic, attempting to be funny.
    - nerdponx 6y ago
      Yeah, people are annoying, with their demands to use software themselves. It would be much easier for everyone if computers were controlled by an elite group of engineers who could hide the complexity from the rest of us. Perhaps they could wear labcoats. What are you even talking about? It sounds like you're upset that cutting-edge technology still requires training & expertise to use and deploy effectively in industry.
      - dang 6y ago
        Hey, can you please not take HN threads further into flamewar? We're trying to avoid that sort of thing here. If a comment contains a swipe, please don't escalate. Also, it's good to check if there's something in your earlier comment that might have been provocative in its own right (which there was: "What the fuck more do you want?" is a hop flameward). https://news.ycombinator.com/newsguidelines.html https://news.ycombinator.com/newsguidelines.html
colincooke 6y ago
There are many many practical example of modern ML (especially DL). Would be interesting to hear why you think those examples are not indicative of a field which is useful/not a hoax.
- whatshisface 6y ago
  Word on the street is the level of superficially attractive papers with no merit is very high in ML literature.
  - colincooke 6y ago
    That is absolutely correct, but is sadly the case in a lot of fields. It doesn't mean that the practical results we see (AlphaFold, Imagenet Performance, NLP performance, Robotic control with RL) isn't amazing progress. Luckily due to so many people using ML these days, what's useful vs. fluff gets sorted out over time.
    - skybrian 6y ago
      Is AlphaFold a practical result? Winning a competition isn't the same as production use. It would be interesting to read about how it's being used. This is a good list of promising work, but showing practicality would need more explanation.
      - alisonkisk 6y ago
        It's a fair question. Is DeepMind famous for its amazingly smart toys because it's useful similarly-smart stuff is secret? Or public but boring? Or doesn't exist? (similar for Boston Dynamics, and IBM Watson)
      - beowulfey 6y ago
        In theory it should be practical. The first generation has been adapted by other teams into excellent prediction servers that can be used now. The second gen is way more hush hush and has yet to be vetted, so we’ll have to see. I am watching for news of it eagerly!
ve55 6y ago
Something can be both legitimately revolutionary/interesting, but also significantly over-hyped and misrepresented, often with strong for-profit incentives. Some recent good examples of this include progress in cryptocurrencies, decentralization, and ML/AI.
- zozbot234 6y ago
  Sure, but many people disagree that ML itself is revolutionary. The basics of it were known (referred to, quite appropriately, as 'data mining') as early as the 1990s and perhaps earlier. We've added a smattering of new techniques since then, and compute power has been expanded via GPGPU, but there was no "revolutionary" shift in the field. Even multi-layer ("deep") neural networks are very old tech.
  - rorykoehler 6y ago
    As evidenced by gpt3 we still don’t know the best way to use deep learning. More data improved the outcome dramatically. What else might surprise us?
  - colincooke 6y ago
    The smattering of new techniques seem to have made the difference between success on toy problems vs. being able to match or exceed human performance on many difficult tasks. So while naysayers are correct that "the math hasn't changed since the 90s!", enough has changed to make calling DL a paradigm shift accurate. For reference, I can now get an intern to images for a few hours, then train a black box algorithm to automate their efforts in another few hours. This algorithm is sensitive, brittle, and may have perfomance issues, but it's still already orders of magnitudes better than what took days or months of effort prior. That to me is a revolution, regardless of the math.
    - tasogare 6y ago
      > then train a black box algorithm This is part of the problem. Finding a the value of a few hyperparameters is hardly something I consider interesting science.
nerdponx 6y ago
You can't have a master's in ML and seriously argue this. It doesn't even make sense, it's like saying marijuana is a hoax because my uncle smokes pot and still got cancer. Here are some alternative statements that make more sense (and contain more truth): * There is a lot of snake oil and outright fraud being sold to unwitting managers. * There is a lot of empty hype being fed to general public through the pop sci media and mainstream news. * Deep learning specifically has not borne fruit in all (edit: or even most) problem domains. * Lack of good quality data (and qualified people to analyze it) is a bigger problem than lack of advanced models and computing power.
- zozbot234 6y ago
  > Lack of good quality data (and qualified people to analyze it) is a bigger problem than lack of advanced models and computing power.' There's plenty of data and compute power, but what's often lacking in the ML field is precisely models that reflect reasonable priors for one's given use case. Good feature engineering (often relying on domain experts) is similarly underrated. You see this again and again when looking at how robust SOTA results are achieved. In a way, this means that good (non-"hoax") ML is ultimately a lot more similar to traditional statistics than most practitioners are willing to acknowledge.
  - nerdponx 6y ago
    Very good point and well-said.
  - teataster 6y ago
    While I wholeheartedly agree with your point, he said good quality data. I am currently working with real estate data. There is no way of knowing whether an entry in the database is a house or a house's floor. I had a project at a death insurance company (they pay your funeral). They had customers dying and coming back to life. You would say those are core business issues that should be dealt with.
nmfisher 6y ago
Have you used speech recognition (or speech synthesis) lately? It's incredible, leaps and bounds ahead of where it was a decade ago. Not everything in ML is as rosy as the papers make it out to be, but to call it a "hoax" is going way too far.
- aborsy 6y ago
  There are a lot of things that look incredible, but don’t constitute much progress scientifically. For example, a rocket landing on moon looks incredible. But I don’t think physicists would consider moon landing notable progress. So, with data and compute you could do applications.
  - btilly 6y ago
    Do you have any idea how many areas of science were opened up by our attempt to get to the Moon? The range is literally from discovering that unit tests are good in software to discovering the Van Allen belts to learning about the geology and history of the Moon from the rocks that we brought back. Could we have learned more science by doing something else with the money? Of course. But it is a dramatic overstatement to say that the Apollo program didn't "constitute any progress scientifically."
    - aborsy 6y ago
      Yes, and I tell you very few if any. If you don’t believe an anonymous person here, see what prominent physicists say clearly on this topic, e.g., Steven Weinberg. This is not to dismiss experimental research which is quite important, but to distinguish (experimental or theoretical) science from product development.
      - btilly 6y ago
        Why don't we see what he says? In https://www.thespacereview.com/article/1037/1 https://www.thespacereview.com/article/1037/1 you'll find that he is very critical of manned spaceflight in general, but about Apollo he says, "No, at the time of Apollo, the astronauts did do some useful things. They brought back Lunar samples. They placed a laser reflector on the Moon that has been used ever since to monitor the motion of the Moon with incredible accuracy." Earlier in the same interview he criticized NASA for canceling Apollo 18 and 19 because he wanted the science that would have been done, to be done. I guess he didn't say what you thought he said. That said, his criticism of NASA's efforts with manned flight isn't because he doesn't think that it is useless to go to the moon. It is because it takes a lot of work to get humans there, and robots can do the job much more safely and cheaper. Which also explains why he thought Apollo was useful. At the time the technology of robotics was much worse so humans were the only way to do the job.
sorenjan 6y ago
I think it depends on what you call it. Instead of calling it AI or even ML you could call it pattern recognition or automatic model parameter estimation, but it doesn't sound as cool.
_carbyau_ 6y ago
Someone else on HN used the phrase "cognitive automation". I like this as a descriptive phrase. ML won't be "intelligence" per se, but it can do repetitive tasks that otherwise required thought.
- fuzzfactor 6y ago
  This is what I was doing in 1980. With emerging digital tools at the time which made it possible to store & retrieve datasets that I was already interpreting in detail anyway. Since it was programmable too, ended up using the memory to store the key points from many permutations of well-characterized raw training data, then running that against new datasets to give me advice on how to save time on the greatly reduced manual work remaining.
mpfundstein 6y ago
so we just discard all the amazing results we had in the last decade? alpha go, alphafold2, gpt-3, image recognition benchmarks etc? i work in ML as well* and while its certainly overhyped and often sold prematurelt, a lot of the stuff is real and works very very well. * i build stuff for cities to more efficiently monitor their citizens (no tracking or personal identification, just aggregate numbers)

timr 6y ago

It isn't a hoax, but the OP is exactly right: if it were more usable, people would see it for what it is, and not for what the silly media narrative makes it sounds like. As long as your technology is only usable by a high priesthood, you can make it look like magic.

nerdponx 6y ago
Why do people feel entitled to "usable" ML at all? In the last 5 years, we have made incomprehensibly huge improvements in power and usability. It's an active field, and improvements are still coming at a steady pace. We have already revolutionized search, natural language processing & machine translation, image/audio/video processing, robotics, game AI, and advertising (for better or worse). And on top of all this, we have significantly reduced the "time to first useful model", and we have significantly lowered the math and programming requirements for building and implementing models. And now we have transfer learning, which lets any old Joe Schmo benefit from massive computing power and datasets to build small on-device models that blow away SOTA accuracy from even a few years ago. Oh, and the ML tooling ecosystem has become a substantial source of innovation in programming language design, "developer UX", and "data ops". What the fuck more do you want? The people who seem the most upset that ML isn't magic seem to be the most confused about what ML even is and does.
- timr 6y ago
  > Why do people feel entitled to "usable" ML at all? Yeah, people are annoying, with their demands to use software themselves. It would be much easier for everyone if computers were controlled by an elite group of engineers who could hide the complexity from the rest of us. Perhaps they could wear labcoats. > What the fuck more do you want? If I knew the answer to that, life would be a lot simpler.
  - dang 6y ago
    > Yeah, people are annoying, with their demands to use software themselves. It would be much easier for everyone if computers were controlled by an elite group of engineers who could hide the complexity from the rest of us. Perhaps they could wear labcoats. That's clearly not what nerdponx meant. Can you please stick to the site guidelines? "Please respond to the strongest plausible interpretation of what someone says, not a weaker one that's easier to criticize. Assume good faith." https://news.ycombinator.com/newsguidelines.html https://news.ycombinator.com/newsguidelines.html > If I knew the answer to that, life would be a lot simpler. That was a nice de-escalation.
    - timr 6y ago
      > That's clearly not what nerdponx meant. I disagree. I made a straightforward interpretation of what was written. Given what came after the part I quoted, you really have to stretch to interpret it differently. The OP underscored the same point using profanity. I made a response that was clearly sardonic, attempting to be funny.
  - nerdponx 6y ago
    Yeah, people are annoying, with their demands to use software themselves. It would be much easier for everyone if computers were controlled by an elite group of engineers who could hide the complexity from the rest of us. Perhaps they could wear labcoats. What are you even talking about? It sounds like you're upset that cutting-edge technology still requires training & expertise to use and deploy effectively in industry.
    - dang 6y ago
      Hey, can you please not take HN threads further into flamewar? We're trying to avoid that sort of thing here. If a comment contains a swipe, please don't escalate. Also, it's good to check if there's something in your earlier comment that might have been provocative in its own right (which there was: "What the fuck more do you want?" is a hop flameward). https://news.ycombinator.com/newsguidelines.html https://news.ycombinator.com/newsguidelines.html

colincooke 6y ago

There are many many practical example of modern ML (especially DL). Would be interesting to hear why you think those examples are not indicative of a field which is useful/not a hoax.

whatshisface 6y ago
Word on the street is the level of superficially attractive papers with no merit is very high in ML literature.
- colincooke 6y ago
  That is absolutely correct, but is sadly the case in a lot of fields. It doesn't mean that the practical results we see (AlphaFold, Imagenet Performance, NLP performance, Robotic control with RL) isn't amazing progress. Luckily due to so many people using ML these days, what's useful vs. fluff gets sorted out over time.
  - skybrian 6y ago
    Is AlphaFold a practical result? Winning a competition isn't the same as production use. It would be interesting to read about how it's being used. This is a good list of promising work, but showing practicality would need more explanation.
    - alisonkisk 6y ago
      It's a fair question. Is DeepMind famous for its amazingly smart toys because it's useful similarly-smart stuff is secret? Or public but boring? Or doesn't exist? (similar for Boston Dynamics, and IBM Watson)
    - beowulfey 6y ago
      In theory it should be practical. The first generation has been adapted by other teams into excellent prediction servers that can be used now. The second gen is way more hush hush and has yet to be vetted, so we’ll have to see. I am watching for news of it eagerly!

ve55 6y ago

Something can be both legitimately revolutionary/interesting, but also significantly over-hyped and misrepresented, often with strong for-profit incentives. Some recent good examples of this include progress in cryptocurrencies, decentralization, and ML/AI.

zozbot234 6y ago
Sure, but many people disagree that ML itself is revolutionary. The basics of it were known (referred to, quite appropriately, as 'data mining') as early as the 1990s and perhaps earlier. We've added a smattering of new techniques since then, and compute power has been expanded via GPGPU, but there was no "revolutionary" shift in the field. Even multi-layer ("deep") neural networks are very old tech.
- rorykoehler 6y ago
  As evidenced by gpt3 we still don’t know the best way to use deep learning. More data improved the outcome dramatically. What else might surprise us?
- colincooke 6y ago
  The smattering of new techniques seem to have made the difference between success on toy problems vs. being able to match or exceed human performance on many difficult tasks. So while naysayers are correct that "the math hasn't changed since the 90s!", enough has changed to make calling DL a paradigm shift accurate. For reference, I can now get an intern to images for a few hours, then train a black box algorithm to automate their efforts in another few hours. This algorithm is sensitive, brittle, and may have perfomance issues, but it's still already orders of magnitudes better than what took days or months of effort prior. That to me is a revolution, regardless of the math.
  - tasogare 6y ago
    > then train a black box algorithm This is part of the problem. Finding a the value of a few hyperparameters is hardly something I consider interesting science.

nerdponx 6y ago

You can't have a master's in ML and seriously argue this. It doesn't even make sense, it's like saying marijuana is a hoax because my uncle smokes pot and still got cancer. Here are some alternative statements that make more sense (and contain more truth): * There is a lot of snake oil and outright fraud being sold to unwitting managers. * There is a lot of empty hype being fed to general public through the pop sci media and mainstream news. * Deep learning specifically has not borne fruit in all (edit: or even most) problem domains. * Lack of good quality data (and qualified people to analyze it) is a bigger problem than lack of advanced models and computing power.

zozbot234 6y ago
> Lack of good quality data (and qualified people to analyze it) is a bigger problem than lack of advanced models and computing power.' There's plenty of data and compute power, but what's often lacking in the ML field is precisely models that reflect reasonable priors for one's given use case. Good feature engineering (often relying on domain experts) is similarly underrated. You see this again and again when looking at how robust SOTA results are achieved. In a way, this means that good (non-"hoax") ML is ultimately a lot more similar to traditional statistics than most practitioners are willing to acknowledge.
- nerdponx 6y ago
  Very good point and well-said.
- teataster 6y ago
  While I wholeheartedly agree with your point, he said good quality data. I am currently working with real estate data. There is no way of knowing whether an entry in the database is a house or a house's floor. I had a project at a death insurance company (they pay your funeral). They had customers dying and coming back to life. You would say those are core business issues that should be dealt with.

nmfisher 6y ago

Have you used speech recognition (or speech synthesis) lately? It's incredible, leaps and bounds ahead of where it was a decade ago. Not everything in ML is as rosy as the papers make it out to be, but to call it a "hoax" is going way too far.

aborsy 6y ago
There are a lot of things that look incredible, but don’t constitute much progress scientifically. For example, a rocket landing on moon looks incredible. But I don’t think physicists would consider moon landing notable progress. So, with data and compute you could do applications.
- btilly 6y ago
  Do you have any idea how many areas of science were opened up by our attempt to get to the Moon? The range is literally from discovering that unit tests are good in software to discovering the Van Allen belts to learning about the geology and history of the Moon from the rocks that we brought back. Could we have learned more science by doing something else with the money? Of course. But it is a dramatic overstatement to say that the Apollo program didn't "constitute any progress scientifically."
  - aborsy 6y ago
    Yes, and I tell you very few if any. If you don’t believe an anonymous person here, see what prominent physicists say clearly on this topic, e.g., Steven Weinberg. This is not to dismiss experimental research which is quite important, but to distinguish (experimental or theoretical) science from product development.
    - btilly 6y ago
      Why don't we see what he says? In https://www.thespacereview.com/article/1037/1 https://www.thespacereview.com/article/1037/1 you'll find that he is very critical of manned spaceflight in general, but about Apollo he says, "No, at the time of Apollo, the astronauts did do some useful things. They brought back Lunar samples. They placed a laser reflector on the Moon that has been used ever since to monitor the motion of the Moon with incredible accuracy." Earlier in the same interview he criticized NASA for canceling Apollo 18 and 19 because he wanted the science that would have been done, to be done. I guess he didn't say what you thought he said. That said, his criticism of NASA's efforts with manned flight isn't because he doesn't think that it is useless to go to the moon. It is because it takes a lot of work to get humans there, and robots can do the job much more safely and cheaper. Which also explains why he thought Apollo was useful. At the time the technology of robotics was much worse so humans were the only way to do the job.

sorenjan 6y ago

I think it depends on what you call it. Instead of calling it AI or even ML you could call it pattern recognition or automatic model parameter estimation, but it doesn't sound as cool.

_carbyau_ 6y ago

Someone else on HN used the phrase "cognitive automation". I like this as a descriptive phrase. ML won't be "intelligence" per se, but it can do repetitive tasks that otherwise required thought.

fuzzfactor 6y ago
This is what I was doing in 1980. With emerging digital tools at the time which made it possible to store & retrieve datasets that I was already interpreting in detail anyway. Since it was programmable too, ended up using the memory to store the key points from many permutations of well-characterized raw training data, then running that against new datasets to give me advice on how to save time on the greatly reduced manual work remaining.

mpfundstein 6y ago

so we just discard all the amazing results we had in the last decade? alpha go, alphafold2, gpt-3, image recognition benchmarks etc? i work in ML as well* and while its certainly overhyped and often sold prematurelt, a lot of the stuff is real and works very very well. * i build stuff for cities to more efficiently monitor their citizens (no tracking or personal identification, just aggregate numbers)