3 ms·

That would cause the inverse problem: the models would overly rely on currently extant animals so the results would look too much like the animals we know. We

by dtech 9y ago

That would cause the inverse problem: the models would overly rely on currently extant animals so the results would look too much like the animals we know. We could probably train a model if we had an accurate training set of dinosaurs, but then it would be pointless.

hwillis 9y ago
I think that must be a fallacy, but I'm not sure which. How can we do worse by including more information? It's not like our pictures today are based on a better source of information. Hell, the article even points out that some of the problem is simply that artists aren't even accurate to the skeletons, the only real information that we really have!
- dsacco 9y ago
  > I think that must be a fallacy, but I'm not sure which. How can we do worse by including more information? I don't think the term "fallacy" applies either way because the parent's point not reducible to a logical error. It's more a matter of perspective. Theoretically speaking, more information is always better. But that's only if you can feasibly work with arbitrary amounts of data. For example, in a deterministic universe we could simulate every instance of time forwards and backwards throughout history, but by definition we'd require more computing power than is possible to perfectly simulate our universe. Similarly, if your information far outstrips your computational capabilities, you can no longer meaningfully work with the data. There are pretty hard economic limits on what kind of data we can work with today, so practically speaking it's accurate to say that more information can be a net negative. In particular, a preoccupation with acquiring more data can lead the analysis to focusing on the wrong data, or it can lead to a hopelessly noisy analysis, or it can lead to subjective decisions about which data to use that meaningfully alters the conclusions. Mathematically speaking, basic results from combinatorics (e.g. Ramsey theory) guarantee that at a sufficiently great size, your data is guaranteed to contain spurious correlations that have to be controlled for. Whenever I'm working on an applied data analysis project I try to use as little information as I can from the outset. Sometimes this means fewer dimensions of data points, or smaller slices of a timeseries. But either way it results in less overall information to process.
- dragonwriter 9y ago
  > How can we do worse by including more information? If the information is not relevant but mistakenly assumed to be, because (e.g.) it is not, but treated as, representative of the class of animals to which it is applied, it can produce a worse model than one constructed without the bad information.
- YeGoblynQueenne 9y ago
  >> How can we do worse by including more information? By including information that is not relevant, or that is redundant. Search online for Feature Subset Selection. It's a thing. Given that more information makes training more expensive, if a lot of your data does not contribute to your model then you have a very good reason to exclude it. Conversely, you have a motive to not include irrelevant information.

by dtech 9y ago

hwillis 9y ago

I think that must be a fallacy, but I'm not sure which. How can we do worse by including more information? It's not like our pictures today are based on a better source of information. Hell, the article even points out that some of the problem is simply that artists aren't even accurate to the skeletons, the only real information that we really have!

dsacco 9y ago
> I think that must be a fallacy, but I'm not sure which. How can we do worse by including more information? I don't think the term "fallacy" applies either way because the parent's point not reducible to a logical error. It's more a matter of perspective. Theoretically speaking, more information is always better. But that's only if you can feasibly work with arbitrary amounts of data. For example, in a deterministic universe we could simulate every instance of time forwards and backwards throughout history, but by definition we'd require more computing power than is possible to perfectly simulate our universe. Similarly, if your information far outstrips your computational capabilities, you can no longer meaningfully work with the data. There are pretty hard economic limits on what kind of data we can work with today, so practically speaking it's accurate to say that more information can be a net negative. In particular, a preoccupation with acquiring more data can lead the analysis to focusing on the wrong data, or it can lead to a hopelessly noisy analysis, or it can lead to subjective decisions about which data to use that meaningfully alters the conclusions. Mathematically speaking, basic results from combinatorics (e.g. Ramsey theory) guarantee that at a sufficiently great size, your data is guaranteed to contain spurious correlations that have to be controlled for. Whenever I'm working on an applied data analysis project I try to use as little information as I can from the outset. Sometimes this means fewer dimensions of data points, or smaller slices of a timeseries. But either way it results in less overall information to process.
dragonwriter 9y ago
> How can we do worse by including more information? If the information is not relevant but mistakenly assumed to be, because (e.g.) it is not, but treated as, representative of the class of animals to which it is applied, it can produce a worse model than one constructed without the bad information.
YeGoblynQueenne 9y ago
>> How can we do worse by including more information? By including information that is not relevant, or that is redundant. Search online for Feature Subset Selection. It's a thing. Given that more information makes training more expensive, if a lot of your data does not contribute to your model then you have a very good reason to exclude it. Conversely, you have a motive to not include irrelevant information.