4 ms·

Not just publicly available. They should be reproducible. I image a scenario of a journalist who wanted to investigate whether some post on a social network's

by pllbnk 5y ago

Not just publicly available. They should be reproducible. I image a scenario of a journalist who wanted to investigate whether some post on a social network's wall was pushed by a group of certain interests or maybe analyze the trends and understand who's behind them. There should be a way to receive a set of inputs and a reproducible way to get the same output. It could start with a user interface where a user could click on a post and they would see the full context of what has led to this post being shown to them at a given moment. I think it would help eliminate the current trend of shoving whatever machine learning models deem the most addictive and would make the web more transparent and user friendly because then the developers would have to optimize differently. So to summarize, I think the algorithms AND datasets should be public. In addition to that, even if they were not public, their inputs and outputs should be reproducible and accessible.

terafo 5y ago
If they publish dataset and code, but all decisions are made by trillion parameter NN, which costs millions to train, is it really reproducible?
- pllbnk 5y ago
  The point is that the social networks should be mandated to make them reproducible. If they can't do it with their ML models (and they can't), then they should come up with some other way. As another person in this thread suggested, they could just display chronologically ordered posts by the accounts users follow or maybe some other useful way. As I said, that would make social networks more transparent and user friendly, just by eliminating the black box ML models.
  - terafo 5y ago
    The thing is that it will make situation neither more transparent nor more user-friendly. You will get expert-systems that are not only so complicated that no one really understand the whole system, but they will provide even worse results(here goes user-friendliness). And as I said in the reply to the person that proposed that everything should be in chronological order, there are inherent reasons why it isn't being used that much anymore(it requires discipline, it doesn't suit short sessions). Problem with social networks is multi-faceted and won't be solved just by mandated openness, reproducibility or direct democracy. Nor it will be by the three of them. And badly-designed regulation is certainly going to make situation even worse.

by pllbnk 5y ago

terafo 5y ago

If they publish dataset and code, but all decisions are made by trillion parameter NN, which costs millions to train, is it really reproducible?

pllbnk 5y ago
The point is that the social networks should be mandated to make them reproducible. If they can't do it with their ML models (and they can't), then they should come up with some other way. As another person in this thread suggested, they could just display chronologically ordered posts by the accounts users follow or maybe some other useful way. As I said, that would make social networks more transparent and user friendly, just by eliminating the black box ML models.
- terafo 5y ago
  The thing is that it will make situation neither more transparent nor more user-friendly. You will get expert-systems that are not only so complicated that no one really understand the whole system, but they will provide even worse results(here goes user-friendliness). And as I said in the reply to the person that proposed that everything should be in chronological order, there are inherent reasons why it isn't being used that much anymore(it requires discipline, it doesn't suit short sessions). Problem with social networks is multi-faceted and won't be solved just by mandated openness, reproducibility or direct democracy. Nor it will be by the three of them. And badly-designed regulation is certainly going to make situation even worse.