> Of course they are there are plenty of cases where the 'of course' does not apply. In computer vision for example models generally have a bias towards scale

> Of course they are there are plenty of cases where the 'of course' does not apply. In computer vision for example models generally have a bias towards scale invariance but struggle with rotational invariance. This is not a problem of 'bad data', but of the kind of features that the architecture is prone to extract and represent while it struggles with others. That's why there is an entire zoo of different ML architectures and systems, because there is no magical uniform, algorithm that performs equally well in every domain. CNN's excel at spatial data, while RNNs perform better on sequential data, and so on. Of course one can attempt to reframe this as a 'data problem' and argue that just means you need to input 100x more data of X than of Y, but that actually just shows that algorithm performance is not uniform, and the more productive thing would actually be to understand strengths and weaknesses of architecture.