4 ms·

There is this famous story about research into image recognition back in 1960s - everyone involved thought it was literally a matter of few months, maybe a year

There is this famous story about research into image recognition back in 1960s - everyone involved thought it was literally a matter of few months, maybe a year at most, to have a system that has 100% accurate recognition capabilities. Yet we're in 2022 and the best of the best image recognition systems spew out absolute nonsense no matter how much money is poured into them. It's just the same sort of issue. It can work amazingly well 90% of the time, but the remaining 10% is basically impossible without some kind of actual general AI that can understand the context of what it's looking at. I honestly believe that current approaches to this tech will not get us to general self driving tech that is safe enough for use on our roads.

dekhn 5y ago
the mark i perceptron did image recognition with a single layer neural network and was surprisingly good at it. Some days I think Rosenblatt was on to something, but about 40 years ahead of the time.
seanmcdirmid 5y ago
Likewise, a lot of computer scientists didn't take Engelbart's 1969 "mother of all demos" seriously because they thought general artificial intelligence was right around the corner. Why bother with a mouse and bitmapped display when you can just tell the computer what you want?
xipho 5y ago
Another story in a similar vein. There are many, many stories about how the advent of DNA sequencing would finally solve the issues of creating the tree of life for all species on Earth, and solve the problem of defining species boundaries, this was some 20 years ago or more. We're still nowhere close thanks largely to the unimaginable quantity of biodiversity out there. We have the low-hanging fruit genomes (specimens readily available, known to be commercially important, etc), but that's about it. Even with those "genomes" (most are not curated to any real level of completion) we still haven't solved compute at scale, even small analyses (10k terminals, 2k genes) take months if not years to comprehensively run (i.e. explore parameter spaces), albeit thanks in part to limits on resources available to scientists. The Earth is big and complex, I can't see the singularity on the horizon yet, but my vision is aging.
- dekhn 5y ago
  I solved the computational problem for genomics already, since it's a mostly-embarassingly-parallel problem (in paticular, I made an idle cycle harvester at Google that applied 1+M xeon cores to science problems, one of which was genomic assembly). I don't know that creating super-accurate species trees is really the most important problem to solve in biology, though. Certainly DNA sequencing for health has been a really mixed bag.
  - xipho 5y ago
    Nice work with that solution! I did similar! I solved the earth's hunger issues, funny thing is we produce enough food, we just can't deploy it to the people who need it. Problem is I can't figure out the logistics of how to make it work... I think maybe you have the same problem? We should compare notes.
    - t_serpico 5y ago
      lol
  - xipho 5y ago
    > Certainly DNA sequencing for health has been a really mixed bag. Ah, the old why bother since it's a "mixed bag" argument. If it wasn't a mixed bag I'd be really worried, as someone has some pretty good snake oil. I vaguely recall something about sequencing and its role in creating a vaciine that turned out to be pretty useful, something recent maybe?
    - dekhn 5y ago
      I don't have a complaint about sequencing that leads to vaccines. That's what sequencing is for. What i'm troubled by is the projects that sequence a million people, using $1B to do so, and then just dump a bunch of genomes with some hand-wavy claims about how it's going to cure cancer. The costs here are rathre significant, but the health outcomes are not.
- rayiner 5y ago
mardifoufs 5y ago
What do you mean spout absolute non sense? The progress in image recognition has been insane in the past decade so your comment might have been accurate 10 years ago. If there's one thing DL/ML are very good at it's computer vision
- gambiting 5y ago
  Sorry I thought it was clear from the rest of my comment. Image recognition is amazing in 90% of the cases, and then in 10% it's worse than a blind human. Another famous example of that is the(now solved) Google image recognition saying with 99% certainty that a sofa in a zebra print is in fact, a zebra(4 legs, zebra pattern = must be a zebra). Look at any demonstrations from Tesla at what comes out of their clasiffier as the car drives around, sure, most of it is correct but some of it is insane, the car thinking there's a ship or a cow in the middle of the road when there's nothing there. Sure the car doesn't act on it, but the fact that the recognition gets it so catastrophically wrong is indicative of the quality of the whole thing.

dekhn 5y ago

the mark i perceptron did image recognition with a single layer neural network and was surprisingly good at it. Some days I think Rosenblatt was on to something, but about 40 years ahead of the time.

seanmcdirmid 5y ago

Likewise, a lot of computer scientists didn't take Engelbart's 1969 "mother of all demos" seriously because they thought general artificial intelligence was right around the corner. Why bother with a mouse and bitmapped display when you can just tell the computer what you want?

xipho 5y ago

Another story in a similar vein. There are many, many stories about how the advent of DNA sequencing would finally solve the issues of creating the tree of life for all species on Earth, and solve the problem of defining species boundaries, this was some 20 years ago or more. We're still nowhere close thanks largely to the unimaginable quantity of biodiversity out there. We have the low-hanging fruit genomes (specimens readily available, known to be commercially important, etc), but that's about it. Even with those "genomes" (most are not curated to any real level of completion) we still haven't solved compute at scale, even small analyses (10k terminals, 2k genes) take months if not years to comprehensively run (i.e. explore parameter spaces), albeit thanks in part to limits on resources available to scientists. The Earth is big and complex, I can't see the singularity on the horizon yet, but my vision is aging.

dekhn 5y ago
I solved the computational problem for genomics already, since it's a mostly-embarassingly-parallel problem (in paticular, I made an idle cycle harvester at Google that applied 1+M xeon cores to science problems, one of which was genomic assembly). I don't know that creating super-accurate species trees is really the most important problem to solve in biology, though. Certainly DNA sequencing for health has been a really mixed bag.
- xipho 5y ago
  Nice work with that solution! I did similar! I solved the earth's hunger issues, funny thing is we produce enough food, we just can't deploy it to the people who need it. Problem is I can't figure out the logistics of how to make it work... I think maybe you have the same problem? We should compare notes.
  - t_serpico 5y ago
    lol
- xipho 5y ago
  > Certainly DNA sequencing for health has been a really mixed bag. Ah, the old why bother since it's a "mixed bag" argument. If it wasn't a mixed bag I'd be really worried, as someone has some pretty good snake oil. I vaguely recall something about sequencing and its role in creating a vaciine that turned out to be pretty useful, something recent maybe?
  - dekhn 5y ago
    I don't have a complaint about sequencing that leads to vaccines. That's what sequencing is for. What i'm troubled by is the projects that sequence a million people, using $1B to do so, and then just dump a bunch of genomes with some hand-wavy claims about how it's going to cure cancer. The costs here are rathre significant, but the health outcomes are not.
rayiner 5y ago

mardifoufs 5y ago

What do you mean spout absolute non sense? The progress in image recognition has been insane in the past decade so your comment might have been accurate 10 years ago. If there's one thing DL/ML are very good at it's computer vision

gambiting 5y ago
Sorry I thought it was clear from the rest of my comment. Image recognition is amazing in 90% of the cases, and then in 10% it's worse than a blind human. Another famous example of that is the(now solved) Google image recognition saying with 99% certainty that a sofa in a zebra print is in fact, a zebra(4 legs, zebra pattern = must be a zebra). Look at any demonstrations from Tesla at what comes out of their clasiffier as the car drives around, sure, most of it is correct but some of it is insane, the car thinking there's a ship or a cow in the middle of the road when there's nothing there. Sure the car doesn't act on it, but the fact that the recognition gets it so catastrophically wrong is indicative of the quality of the whole thing.