Hacker News Reader

Comment by snek_case | Hacker News Reader

There's this dogmatic idea in the machine learning field that only real world data is valuable.

Data augmentation is a thing within machine learning (deep learning) so that's a big generalization.

Data augmentation is a big thing in applied machine learning. Deep learning practitioners use data augmentation because it works incredibly well. However, deep learning researchers tend to view data augmentation as some kind of dirty trick that wouldn't be necessary if you just had a bigger dataset.

It's also about the issue that people will bias against generating the "difficult" data, even if only subconsciously. Real-world validation is essential to reveal whether you have actually worked through the full phase space.

Depends. As others have said, you can be intentionally adversarial (looking for corner cases) when testing your system, in a way that real data isn't.

I guess the field is constantly making a decision between 'do we want to outperform humans, but lose interpretability' and 'we should always be interpretable'

Human-generated artificial data will always contain the human's assumptions. But this data might not contain the 'superhuman' element that leads to the ML-system outperforming humans, because we don't know what that is (yet). Receipts are a good example for artificial data, because they're human-made.

But a lot of what we do with ML systems is using data that doesn't come from humans, images of wildlife, satellite images, biological data, etc

This is simply not true. Synthetic data is a huge research arm of ML work.

Reply on news.ycombinator.com