Do you think deep learning models could create this more accurately? The training set would be known skeletal structures to fleshed out representations.
Although it's perfectly possible that dinosaurs were effectively massive birds of paradise with ridiculous mating accoutrements...
However, taking the hippo as an example, some convex bone shapes seem to be a hint for neighboring body fat and muscle. I'd check first, if current artists would do better after training themselves on living animals. Also, I'd check whether reptiles can get as fat as mammals, etc.
You could probably get decent results just with models/pictures of skeletons and live animals, but I think MRI scans would be ideal- the model could learn to extrapolate actual structures of different tissues directly, and you could model things that don't exist in combination today. Birdlike musculature, body fat and skin over complex cartilaginous structures, with muscle traits of more cold-blooded animals.
There are many zoos with MRI machines, even extra-large ones used for large animals. I wonder how much of that dataset could be assembled? You don't even need very high resolution, centimeters would be plenty for something the size of a horse. Millimeters would be perfect for anything bigger than a sparrow!
[1]: https://upload.wikimedia.org/wikipedia/commons/7/71/Hippo_sk...
We could probably train a model if we had an accurate training set of dinosaurs, but then it would be pointless.
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.
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.
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.