6 ms·

Confusing headline, it seems to imply that all algorithms are going to die, like we'll just stop using algorithms. Obviously that doesn't make sense, but that i

Confusing headline, it seems to imply that all algorithms are going to die, like we'll just stop using algorithms. Obviously that doesn't make sense, but that is what it sounds like. I also question whether we are talking about algorithms, or the data set they are working with or the model created from that. I can forgive confusing an algorithm with its implementation (e.g. the source code), but this goes beyond that.

IAmEveryone 5y ago
A trained model, which is obviously the subject here, is an algorithm.
- robbrown451 5y ago
  I generally wouldn't define algorithm as such, although I can see where the lines between them can become blurry.
  - kaibee 5y ago
    > I generally wouldn't define algorithm as such, although I can see where the lines between them can become blurry. Its pretty clear cut tbh. An algorithm is a set of steps to follow to produce some output. A trained model is, 'hey do these matrix multiplications with these coefficients to get an output'. The fact that the exact coefficients were arrived at via backprop, doesn't make it not an algorithm.
    - The_rationalist 5y ago
      Oh yes so chromium and minecraft are algorithms now. What a useful definition we have here.. Obviously colloquially when people says algorithm, it is an implicature that they mean a hard computing (not soft) serie of steps that achieve a specific goal. In other words, a statistical algorithm does not point to the same category as a deterministic algorithm and an algorithm refer to the later class by default.
      - SirSavary 5y ago
        No reasonable person would define Chromium or Minecraft as algorithms; that's not what the poster was saying.
        The_rationalist 5y ago
        oh yes chromium is not a "set of steps to follow to produce some output" then?.. If the author wasn't meaning what he said, what is the actual criteria he was meaning?
      - mannykannot 5y ago
        > What a useful definition we have here.. Indeed it is - for one thing, it allows us to see that various useful theorems and results about algorithms and computability apply as much to large programs as to small ones, such as the fact that there's no fundamental impediment to porting them between computers with different instruction sets, or running them in virtual machines. What's not so well or usefully defined here is your distinction between hard and soft computing. > In other words, a statistical algorithm does not point to the same category as a deterministic algorithm and an algorithm refer to the later class by default. You appear to be under the misapprehension that the set of statistical algorithms is disjoint from that of deterministic algorithms. I strongly suspect that all the algorithms covered by the article are both statistical in terms of what they compute and deterministic in terms of how they do it.
    - mattmcknight 5y ago
      Not necessarily, as the trained model can just be the matrix. It's more like data, as presumably the same algorithm with different weights, trained by different data would be permissible.
      - riking 5y ago
        ML training is an algorithm that produces an algorithm as output
        mattmcknight 5y ago
        It doesn't really though. The prediction algorithm already existed before training, it just is a particular collection of parameters that the training produces. You wouldn't say changing the interest rate changes the algorithm for calculating interest payments. The weights can be viewed another input to the algorithm. You aren't going to outlaw linear regression, but a particular set of coefficients.
qchris 5y ago
I believe the answer to your second line is "yes." This FTC enforcement action appears to address all of the above: any data that is collected in an illegal manner (in this case, which violates COPPA) and any machine learning models that are created through the use of this data (in their entirety, since you can't un-train a piece of information from a model). Machine learning is somewhat unique in the software world in that the actual useful artifacts are not necessarily strongly tied to the source code itself. You can have the identical source code running at two different companies, but by supplying them with two different training sets, you'll end up with very different outputs. That's what algorithmic destruction is targeted at--not even necessarily the source code or algorithm in the technical sense (you can't destroy "KMeans" or "convolution" as a concept, obviously), but both the data and the model weights that are produced through the use of that data that are used in perform a business action. Those weights are typically stored separately from the source code, and can be extremely expensive to re-create from scratch.
armchairhacker 5y ago
The articles use of “algorithm” is incredibly bad. > The FTC’s new enforcement weapon spells death for algorithms makes it sound like running quick-sort is a federal crime

IAmEveryone 5y ago

A trained model, which is obviously the subject here, is an algorithm.

robbrown451 5y ago
I generally wouldn't define algorithm as such, although I can see where the lines between them can become blurry.
- kaibee 5y ago
  > I generally wouldn't define algorithm as such, although I can see where the lines between them can become blurry. Its pretty clear cut tbh. An algorithm is a set of steps to follow to produce some output. A trained model is, 'hey do these matrix multiplications with these coefficients to get an output'. The fact that the exact coefficients were arrived at via backprop, doesn't make it not an algorithm.
  - The_rationalist 5y ago
    Oh yes so chromium and minecraft are algorithms now. What a useful definition we have here.. Obviously colloquially when people says algorithm, it is an implicature that they mean a hard computing (not soft) serie of steps that achieve a specific goal. In other words, a statistical algorithm does not point to the same category as a deterministic algorithm and an algorithm refer to the later class by default.
    - SirSavary 5y ago
      No reasonable person would define Chromium or Minecraft as algorithms; that's not what the poster was saying.
      - The_rationalist 5y ago
        oh yes chromium is not a "set of steps to follow to produce some output" then?.. If the author wasn't meaning what he said, what is the actual criteria he was meaning?
    - mannykannot 5y ago
      > What a useful definition we have here.. Indeed it is - for one thing, it allows us to see that various useful theorems and results about algorithms and computability apply as much to large programs as to small ones, such as the fact that there's no fundamental impediment to porting them between computers with different instruction sets, or running them in virtual machines. What's not so well or usefully defined here is your distinction between hard and soft computing. > In other words, a statistical algorithm does not point to the same category as a deterministic algorithm and an algorithm refer to the later class by default. You appear to be under the misapprehension that the set of statistical algorithms is disjoint from that of deterministic algorithms. I strongly suspect that all the algorithms covered by the article are both statistical in terms of what they compute and deterministic in terms of how they do it.
  - mattmcknight 5y ago
    Not necessarily, as the trained model can just be the matrix. It's more like data, as presumably the same algorithm with different weights, trained by different data would be permissible.
    - riking 5y ago
      ML training is an algorithm that produces an algorithm as output
      - mattmcknight 5y ago
        It doesn't really though. The prediction algorithm already existed before training, it just is a particular collection of parameters that the training produces. You wouldn't say changing the interest rate changes the algorithm for calculating interest payments. The weights can be viewed another input to the algorithm. You aren't going to outlaw linear regression, but a particular set of coefficients.

qchris 5y ago

I believe the answer to your second line is "yes." This FTC enforcement action appears to address all of the above: any data that is collected in an illegal manner (in this case, which violates COPPA) and any machine learning models that are created through the use of this data (in their entirety, since you can't un-train a piece of information from a model). Machine learning is somewhat unique in the software world in that the actual useful artifacts are not necessarily strongly tied to the source code itself. You can have the identical source code running at two different companies, but by supplying them with two different training sets, you'll end up with very different outputs. That's what algorithmic destruction is targeted at--not even necessarily the source code or algorithm in the technical sense (you can't destroy "KMeans" or "convolution" as a concept, obviously), but both the data and the model weights that are produced through the use of that data that are used in perform a business action. Those weights are typically stored separately from the source code, and can be extremely expensive to re-create from scratch.

armchairhacker 5y ago

The articles use of “algorithm” is incredibly bad. > The FTC’s new enforcement weapon spells death for algorithms makes it sound like running quick-sort is a federal crime