I'm not sure if this is what the article is trying to get at, but here's a concern I have about models vs data: algorithm choice and hyperparameters. I'm no ex

I'm not sure if this is what the article is trying to get at, but here's a concern I have about models vs data: algorithm choice and hyperparameters. I'm no expert but I've trained a few models, and I think there's a fair bit of manual tuning that goes on to maximize accuracy. So in terms of the model vs data dichotomy, could we be optimizing for models based on biased data now, eg lack of labeled black faces in image recognition, and later once we fix the data, the model will still underperform on black faces? I did a simple transfer learning project based on a model that had trained on an enormous data set, and my understanding is that it worked because the lower layers had "figured out" how to look at edges and contrast and such. But could those lower primitive features be biased so that even with more, new, training data, the model won't work so well? (e.g. focusing on contrast, which might not work as well on darker faces). I think this all depends on a sort of "build a model once and retrain on new datasets" approach. Is that how it works in practice? Or is the model re-tuned, architected, etc each time there's a change to the training data? In that case since the model is effectively tied 1 to 1 to the data, I'm not sure it makes much sense to draw this distinction between model and data, right?